WebApi: every operation carries WebIDL's property attributes, not ECMA-262's

[lahma]

Closes #3239, on the terms settled in its adjudication comment: a deliberate per-declaration sweep with an explicit decision for the two singletons, not a blanket generator default.

WebIDL §3.7.6 gives a regular operation { writable, enumerable, configurable }. Jint's web-API operations carried ECMA-262's rule for a built-in function instead — not by choice, but because JsFunctionAttribute had no Flags property until #3238.

The census was wrong twice, and this is the corrected one

The issue said 188 operations of 193. grep counted three hits inside doc comments: there are 190 real declarations, so the sweep is 183.

member kind	total	already right	changed	exempt
Operation ([JsFunction])	190	5	183	2 (scheduler)
Attribute ([JsAccessor])	187	184	2	1 (navigator.userAgent)
Constant, constructor, @@toStringTag, @@iterator	207	207	0	—

185 members across 46 files. Largest movers: console 18, SubtleCrypto.prototype 12, URLSearchParams.prototype 12, Headers/FormData 10 each. Static operations moved too — Response.error/json/redirect, AbortSignal.abort/any/timeout, URL.canParse/parse, ReadableStream.from.

Node 24 agrees member for member wherever it implements the interface. One deviation reported honestly rather than glossed: Object.keys(AbortSignal) is [] in Node (hand-written internals; its WebIDL-generated bindings do follow the rule), and Node omits Event.prototype's constants entirely.

scheduler and navigator keep ECMAScript's attributes

Both are interface instances with no interface object and no interface prototype object, so their members are own properties of the singleton. WebIDL's rule presumes the member sits on the prototype, where it is invisible from the instance.

Node is decisive: Object.keys(navigator) → [], Reflect.ownKeys(navigator) → [], and userAgent is an enumerable accessor on Navigator.prototype. Flipping Jint's would answer ["userAgent"] and ["postTask","yield"] — strictly less browser-accurate. console is the counter-example proving the rule: a WebIDL namespace, whose members genuinely are own properties, so its 18 operations did flip.

The reasoning now lives in the two classes' doc comments, which also lose a stale claim — that they carry "the same pair console, crypto and performance carry", false since #3250 and #3266. The real fix is filed as #3270, which is also what would close the __proto__ containment hazard #3266 identified for these two.

The structural test is the artefact that keeps this true

WebIdlPropertyAttributeTests builds an engine with every WebApiFeatures flag plus a refusing WorkerProvider, diffs the global against a bare engine's, and walks by CLR namespace — 131 objects, 816 members — classifying each by WebIDL member kind and holding it to that kind's attributes. Four facts:

• the sweep itself;
• every generated host is reached — found by reflecting for CreateProperties_Generated, so the list cannot go stale; 79 of 81, with the two worker global scopes recorded as unreachable and re-checked in both directions; plus a checked > 500 non-vacuity floor;
• every recorded divergence is real — the WPT exclusion-table discipline, so a later fix cannot leave a stale exemption behind;
• the two singletons keep empty own keys — the property the divergences rest on.

Adding a web-API member needs no edit here. The class doc is a table of kind → attributes → declaration spelling with a spec anchor each. Only an iterable declaration (4 seeds today) or a deliberate divergence (3 today) ever adds a line.

Verified against unfixed code: 185 failures, exactly the 183 + 2 that changed.

Nine existing tests updated, not the seven the census predicted

Seven in Jint.Tests — CryptoTests, SubtleCryptoTests ×2, SubtleCryptoAesWrapTests, PerformanceTests, EncodingTests, StorageTests (whose for...in now yields one,two,clear,getItem,key,removeItem,setItem,length, the same six prototype names Node sees).

And two the census missed, both in Jint.Tests.PublicInterface: WebApiSubtleCryptoTests.IsAReadOnlyNonEnumerableAttribute, and WebApiEventTests.GivesEveryMemberTheAttributesWebIdlAsksFor — which had pinned the old answer with "Pinned so that changing it is a deliberate act." This is that deliberate act.

Twenty host-class doc paragraphs claiming "one documented simplification: the operations are non-enumerable" are removed, along with the matching clause in the WPT triage record.

Cost: 0 B, bit-identical

1000 iterations after 200 warmup, three rounds each side, identical to the decimal:

	before	after
new Engine()	9,256.0 B	9,256.0 B
new Engine(o => o.UseWebApis())	24,280.0 B	24,280.0 B
…+ 15 web-API property tables materialized	60,912.0 B	60,912.0 B

The mechanism was re-verified rather than taken on trust: HasNoEnumerableOwnStringKeys() is virtual → false, overridden only by ArrayInstance, JsObject and ObjectPrototype, so FixedSlotsAllNonEnumerable is unreachable for every Jint/WebApi/ type, and its only consumer already answered false for these.

Verification

Zero files under Jint/Native/ are touched — and test262 proves it: 102,495 passed / 0 failed on both this branch and an upstream/main baseline on the same machine. WPT 422/422 both ways, not one row moved, exclusion table untouched.

Solution builds Release across all TFMs including net462/netstandard2.x. Jint.Tests 10,514 (net10.0) / 7,256 (net472); Jint.Tests.PublicInterface 2,493 / 1,899; both repeated with JINT_HOST_CONTRACT_VERIFICATION=1.

One caveat reported rather than smoothed: of five test262 runs, four were clean and one intermediate run reported 2 failures whose identity was not captured, with the two runs immediately after clean. Two other agents were running. The change cannot reach test262 — it enables no web APIs and Jint/WebApi/ is unreferenced there — so this is the known load flake, but it is an observation and not nothing.

Two follow-ups filed

• Navigator and Scheduler have no interface object, which is why two of their members cannot carry WebIDL's attributes #3270 — give Navigator and Scheduler real interface objects, which deletes all three divergence entries and closes the containment hazard.
• Web-API prototype members enumerate in name order, where WebIDL specifies attributes-then-operations-then-constants in IDL order #3271 — prototype member enumeration order, newly observable because of this change: Object.keys(Storage.prototype) is ["clear","getItem","key","removeItem","setItem","length"] here against ["length","clear",…] in Node, because WebIDL defines attributes → operations → iteration methods → constants in IDL order while the generator sorts by name. Membership matches; order does not; no corpus row asserts it.

JsFunctionAttribute's doc was deliberately not expanded — it lives in a source-generator raw string reproduced in 54 Verify snapshots, so a 17-line addition would have been ~1,200 lines of duplicated comment churn dominating the diff. The guidance went to AGENTS.md as a fourth Web APIs convention instead.

🤖 Generated with Claude Code