Update test262 suite and implement Iterator.prototype.join

[lahma]

Bumps the pinned test262 suite to defaaf1571cd13b183e3f505c6a06e8db316e593. Three things came in with it, and all three needed engine work to stay green — which is why they are one PR: the SHA bump cannot land without any of them.

Iterator.prototype.join (proposal-iterator-join) is the new feature, with 18 tests. Joining is the easy half; which failures close the receiver is the part worth reading the tests for. Coercing the separator closes it, and so does coercing a value the iterator produced. An abrupt next lookup, a throwing next(), and a protocol violation (a step result that is not an object) all propagate without closing, because IteratorStepValue marks the record done first. Exhaustion does not close either. The separator is coerced before next is read, so a receiver whose next getter throws never gets that far, and next is read exactly once however many steps follow. Nullish values format as the empty string, as they do in Array.prototype.join.

Intl.Locale.prototype.getCollations picked up a normative change (ecma402#1072): a tag matching no available Collator locale no longer falls back to the default locale — it reports the hardcoded root collations « "emoji", "eor" » — and the result is now sorted in lexicographic code unit order. Jint reported a single-element ["default"] for every locale, and "default" is Jint's internal placeholder for "no collation was requested", not an identifier a locale can report at all. It now reports the root collations plus whatever the language adds, sorted; an explicitly requested collation, however it was requested, is still the whole answer. The data comes from the table CollatorConstructor already kept, so nothing new was invented:

tag	before	after
und, und-Latn-US, qtz-CN, tr	["default"]	["emoji","eor"]
de	["default"]	["emoji","eor","phonebk"]
ko	["default"]	["emoji","eor","searchjl","unihan"]
de-u-co-phonebk	["phonebk"]	["phonebk"]

import-defer grew IsModuleSCCEvaluated, and Jint needed it. A module that finished its own body is only really done once the strongly connected component it belongs to is: a member of an async cycle reaches EVALUATED as soon as its body returns, while the cycle root sits in EVALUATING-ASYNC awaiting a top-level await. Reading the member's own status is what GatherAsynchronousTransitiveDependencies and ReadyForSyncExecution were doing, so a deferred subgraph reaching into such a cycle looked free of async dependencies — and the deferring module ran its body, and took its deferred namespace, before the cycle had finished. The new test (import-defer/evaluation-top-level-await/async-cycle-dependency-of-deferred-module) walks exactly that graph, and Jint produced

B, A-before-await, Middle-before-nsD.z, Middle-after-nsD.z, resolve-blocker, C, A-after-await, D

where the spec wants

B, A-before-await, resolve-blocker, A-after-await, Middle-before-nsD.z, D, Middle-after-nsD.z, C

ReadyForSyncExecution refuses an EVALUATED-but-not-SCC-evaluated module rather than asserting on it: the spec asserts LINKED at that point, having convinced itself the state is unreachable, and refusing is both the conservative answer and the one EvaluateSync already turns into the spec's TypeError.

Verification

All in Release, on the rebased branch:

suite	result
Jint.Tests.Test262	99484 passed, 0 failed, 157 skipped
Jint.Tests (net10.0 / net472)	4636 / 4555 passed, 0 failed
Jint.Tests.PublicInterface (net10.0 / net472)	1240 / 1239 passed, 0 failed
both of the above with JINT_HOST_CONTRACT_VERIFICATION=1	0 failed

New in-repo tests cover join's results, its nullish formatting, the separator-before-next ordering, which failures close the receiver and which do not, and the non-object receiver rejection; plus getCollations for the root fallback, a matched locale, an explicit collation, and the "never reports standard/search, always sorted" invariant.

🤖 Generated with Claude Code