javascript: a test that never loads the change is not selected; a mocha load-time throw selects the whole run

[swapnilpaliwal-sd]

JavaScript test selection: a test that never loads the change is not selected, a mocha load-time throw selects the whole run, and the import facts behind both read require('..') and a package's own name.

What changes

impact

• imports_fact resolves a directory import (require('..'), require('../'), a folder with a package.json main/exports), the repository's own package name, a side-effect import './x' and a literal import('./x'); a package whose entry is a build output the checkout lacks names its src/ (or lib/) entry. Before, the root normalised to . and was tried as ./index.js, so most CommonJS test suites imported nothing here and loads_change was false for over a thousand failing test files on one subject.
• loads_change asks whether the test loads the asked declaration's file, not the end of the shortest route (that route can run through another test file's subclass). It says nothing where the target file has no importer, the test imports nothing here, or a computed-name loader (require(path.join(dir, f))) sits in its import closure (imports_dynamic).
• A JavaScript test whose route is in-process (sound, one of a set, dispatch, registered, defines, fixture) and whose file never loads the changed file goes to not_loading_tests and a not counted: line instead of the selection. Key joins, remote hops and spawned scripts are unaffected (the existing key-join case keeps its test).
• load_run (impact.dl): test code that reaches the change through calls alone (a resolved call or one of a resolved set, no defines, no hand-over). A test credited through its module body or describe() block that is one of these carries at_load.

path — from a JavaScript test file that never loads the target's file: no route: X never loads Y, exit 1.

test-impact

• A selected JavaScript file no runner collects is replaced by the collected test files that require it, or moved to not_collected when none does (kept when a test file loads modules by a computed name).
• A mocha run whose specs reach the change at load time lists every spec it collects: mocha loads them in one process and a throw there fails all of them. jest and vitest are per-file and unchanged.
• A function constructor whose header line changes is a TYPE to changed; when impact finds no type at that line it is asked again as the function (before: "could not be resolved to a graph symbol", 0 tests).

Measured

8 JavaScript libraries (5 tuning, 3 held-out) from the 18-repo mutation corpus, cached truth (test files that newly fail when a function throws on entry), the archived no-library protocol; base reproduces the archived numbers exactly. Micro recall / precision:

verb	split	base R	base P	this R	this P
impact --tests	tuning	0.487	0.273	0.511	0.271
impact --tests	held-out	0.735	0.156	0.735	0.478
tests (mutant applied)	tuning	0.455	0.264	0.531	0.284
tests (mutant applied)	held-out	0.735	0.264	0.735	0.518
path test -> target	tuning	0.660	~0.245	0.660	~0.245
path test -> target	held-out	0.658	~0.164	0.658	~0.578

No failing (test file, target) pair is lost by impact, tests or path on any subject; every change in recall is a gain. Path precision is estimated on a fixed sample of the base engine's non-failing selections.

With the runtime dependencies staged on the web-framework subject: tests R 0.460 -> 0.841, P 0.607 -> 0.706.

Cost: cold-solve impact median unchanged on the three largest subjects (0.46/0.46 s, 0.50/0.48 s, 0.31/0.31 s); tests within ±10%.

Tests

• tests/cases/javascript/a-test-that-never-loads-the-change: 3 checks fail on base, pass here; 4 controls (computed-name loader, require('..') / '../' / '../index.js', self-name import, the route of the test that does load the plugin) pass on both.
• tests/cases/javascript/mocha-load-time-fails-the-whole-run: 3 checks fail on base, pass here; controls: a test-body-only change selects its spec alone, and a jest package is not widened.
• python3 tests/run.py (all five languages): 1607 of 1607 checks pass.

Not addressed

• Precision on hubs every test drives (a CLI's parse() reaching help and error paths behind flags): per-test-file coverage shows the false pairs never execute the target; no static signal without path conditions.
• A test helper that calls its own module through module.exports(...), and a parameter defaulted as app = app || new App(), leave one held-out subject's request/response tests unreached (one repo; not tuned on held-out).