Comparing changes

…s) (#25)

…26) * test: add structuredClone suite (opt-in via runStructuredCloneTests) Covers the WHATWG structuredClone global: primitives (including -0, NaN and BigInt), plain objects and arrays, Map/Set/Date/RegExp, the wrapper objects, Error, ArrayBuffer, every typed array and DataView with their offsets, graph identity and cycles, getter invocation, and the uncloneable cases. Transfer coverage asserts both halves of the hand-off: the clone is usable and the source is detached, for a bare buffer, a buffer reached through a typed array, and buffers not present in the cloned value. Rejections cover duplicate entries, non-transferable entries, already detached buffers, and non-iterable transfer lists. Failures are asserted by `.name === "DataCloneError"` rather than `instanceof DOMException`, so the suite runs on runtimes that have no DOMException. The one runtime-specific expectation (a native object is not cloneable) is behind the existing isV8iOS guard. Deliberately kept out of runAllTests(): Android consumes master and does not ship the global yet, so each runtime opts in once it does. * test: pin single capture of the transfer iterator next method WebIDL builds the iterator record once, capturing `next` at creation. The spec drives iteration with an iterator whose `next` is an accessor that yields a working function on the first read and throws on any later one, so an implementation that re-reads `next` per step fails instead of silently diverging. * test: cover SharedArrayBuffer sharing and worker message transfer structuredClone and worker postMessage run on one serialization core, so the transfer rules are asserted from both entry points: a buffer posted with a transfer list arrives usable in the worker while the sender's copy is detached, and duplicate, non-ArrayBuffer and already-detached entries are rejected with the same DataCloneError name as on the structuredClone side. postMessage takes a plain array only, so a non-array transfer list is a TypeError. The SharedArrayBuffer specs feature-detect the constructor and assert sharing in both directions -- written through the clone, read through the original, and back -- plus that one shared buffer stays shared across two references and that it cannot be transferred. The worker specs live here rather than in Workers/index.js because this suite is opt-in: runAllTests() runs on runtimes that have not implemented postMessage transfer yet. Also pins the deliberate asymmetry in host-object handling: postMessage delivers a posted native object as an empty object where structuredClone rejects it. Workers/index.js already pins that posting one does not throw; this adds what the receiver actually sees. * test: gate the shared suite on structuredClone presence The suite was kept out of runAllTests() and wired in per runtime, which meant every runtime that shipped the API had to remember to opt in, and master could not carry the suite at all. It now checks for the global itself: where structuredClone is missing it registers one pending spec and returns, so the suite can live in runAllTests() everywhere and a runtime without the API reports a visible skip instead of a wall of failures. The named export stays, since it is still the way to run just this suite. A gate like this can hide the very regression the suite exists to catch, so the header points implementing runtimes at the unguarded canary they are expected to keep in their own tests.

…exists The detail specs now follow the spec contract on a runtime that ships structuredClone — a distinct snapshot, immune to later mutation, with an uncloneable detail rejected by a DataCloneError-named error — and keep the by-reference assertions as the documented fallback for a runtime that implements the Performance API first.

… bound Reconstructing Date.now() from timeOrigin + performance.now() races three clock reads in one expression; a ~20ms scheduler or GC stall between them on a contended CI host blew the previous 10ms budget (observed as tightly clustered first-attempt failures that pass on rerun). A stall does not repeat across every sample, so the minimum offset over ten samples filters it out, while a genuine anchoring or unit error persists through all of them — 250ms still catches those. Applies to the main-isolate spec and the worker's inner sample alike.

Navigation Menu

Uh oh!

Uh oh!

Comparing changes

Open a pull request

Uh oh!

Commits on Aug 11, 2026

This comparison is taking too long to generate.

Uh oh!