Native parser public API and integration work

The current CSS target is LightPanda-style renderless text. This supersedes the former full-computed-CSS plan and its V0-V9/R3-R7/C6 grammar obligations. Removing these features is an intentional behavior change, not deferred implementation debt. Historical implementation measurements in git history do not describe the current engine.

CSS boundary

Area	Disposition
Syntax, selectors, media	Retained, including bounded work and existing selector specificity.
Style/media/supports/layer/import/font-face CSSOM	Retained; declarations and font descriptors store text.
Other at-rules	Generic CSSRule with source cssText; no cascade effects or specialized interfaces.
V0-V9 typed grammars, math, colors, transforms, all resets	Removed by design: parsing cost is not justified by renderless automation. Catalog names are data, not grammar obligations.
Keyframes, property registrations, container queries, pages/counters/namespaces	Semantic models removed by design; opaque rule text remains.
Computed values	Text cascade, inheritance and small defaults; synthetic width/height only. No unit/color computation.
Variables	Inherited custom text, with optional depth/size-bounded textual substitution in ordinary properties; no typed substitution graph.
Retained limitations	Nested conditional rules/interleaved declarations and import layer()/supports() still have explicit unsupported boundaries; stylesheet byte/MIME work remains below.

UA/shadow/import handling, broader selectors/media conditions, basic synthetic flex geometry and textual variables are deliberate conveniences beyond LightPanda. CSS parsing remains on demand. Public XPath, HTML/XML serialization, syntax-only ParseCss and the native parser improvements remain unchanged. No HTML tokenizer, DOM, XML or XPath work is part of this CSS reduction.

Retained completed slices

• Owned public XPath: NativeXPath.Compile, Evaluate and Select, for native nodes, attributes and namespace contexts; private prepared BCL expressions, typed immutable results, detached-following guard, namespace freshness, cancellation and identity preservation. Evidence: NativeXPathEvaluationTests, NativeXPathNamespaceContextTests, NativeXPathIdentifierTests, public API snapshots and the unsigned packed consumer.
• Public HTML/XML serialization: all five MarkupSerializer routes, immutable options, output-limit taxonomy, cancellation, template/namespace behavior and native shadow selection. Element.AttachShadow, OpenShadowRoot and readonly root metadata provide real public identities; registry identities, Browser attachment policy and mutable parser hooks stay internal. Evidence: MarkupSerializerTests, the existing scalar/tree/shadow/work serializer fixtures, public API snapshots and the unsigned packed consumer.
• Existing whole-sheet syntax API reconciliation: ParseCss and CssStyleSheetSyntax are included in the API snapshots and exercised by the packed consumer.
• Native acceptance corrections: synchronous mutation completion for ARIA reflection, dataset and Selection; CDATASection branding; DOM ancestor/reference-child exception precedence; namespace-wildcard attribute-name casing; native inherited :lang() and Browser-backed :dir(). Language matching does not yet include document/HTTP language metadata fallback.
• Stylesheet cache acceptance: reuse when unchanged and invalidation on child sheet edits, import insertion, resource arrival and import removal. No timing limit was raised.
• XML no-fetch and canonical-output acceptance: all 101 pending reviews are resolved, including the internal-parameter-entity case with no external-resource indication. Every eligible canonical output is compared. DTD processing instructions have immutable public parse metadata, separate from DOM children, with exact data, order and original-input offsets. Evidence: XmlConformanceTests, XmlCorpusTests, XmlDtdProcessingInstructionTests, both public API snapshots and the unsigned packed consumer.
• Parser insertion reactions: native parsing yields after potentially custom element insertion, before processing children, including document.write and reconstructed customized formatting. Browser drains reactions outside the entered parser; nested writes retain the active insertion point. Synchronous construction before attributes and constructor-failure semantics remain open. Evidence: HtmlScriptHandoffTests, CustomElementUpgradeTests and the pinned custom-element WPT cases.
• Cascade layers: named/dotted/anonymous/nested ordering, reversed important precedence, ordinary-property rollback, CSSOM identity/mutation and import prologue handling remain. Custom property rollback keywords now remain text. Evidence: CssLayerRuleTests, NativeCssLayerTests.
• Selector observation scaling: one mutation witness per document rather than candidate; mutation and detached-node adoption still invalidate reads. Deterministic witness-count and invalidation tests cover this. Evidence: SelectorInteractionWorkTests, SelectorControlFactsTests, SwaggerFixtureTests.
• Shadow stylesheet host selectors: :host/:host(...), featureless-host restrictions, predicates/combinators, specificity and encapsulation precedence, with live class/sheet invalidation. DOM queries do not acquire stylesheet host context. Evidence: SelectorHostTests, CssSupportsTests, NativeCssHostTests. :host-context() remains unsupported.

Browser loading, selectors and parser integration

• LOAD-01 - Stylesheet byte decoding. Trace the current native stylesheet loading and import paths, then implement the specified BOM/transport/@charset/environment encoding choice. Done when a local-server fixture checks conflicting/absent labels, malformed sequences, imported-sheet encoding and cancellation; decoded-string CSS parsing remains unchanged. Progress: the parser owns CSS Syntax §3.2 fallback selection (CssStyleSheetEncoding) and drops @charset as a rule; Browser decodes <link> and @import bytes through the engine's WHATWG decoders with BOM override and parent-encoding inheritance (CssStyleSheetDecoding). Evidence: CssStyleSheetEncodingTests, StyleSheetBytesAreDecodedWith…. Malformed-sequence and cancellation fixtures remain.
• LOAD-02 - Stylesheet eligibility and load ordering. Implement MIME/nosniff and relevant response/origin eligibility at the loader boundary. Preserve URL/base metadata and distinguish rejected sheets from empty valid sheets. Done when local-server tests prove load/error events, failed imports, script/style ordering and cache invalidation without altering unrelated fetch policy. Coordinate with LOAD-01.
• SEL-01 - Standalone intrinsic control facts. Provide native producers for the default- submit, placeholder, editability, validity, range and directionality facts currently requiring ISelectorControlFactsFactory. Reuse native form/control state rather than Browser wrappers. Done when standalone matching works across input types, form owners, disabled/read-only states, mutation/adoption and cancellation, while absent interaction state means no focus/hover/active.
• SEL-02 - Host-context and stylesheet pseudo-element audit. Implement :host-context() with shadow-including ancestor matching and normal-context arguments in the existing iterative VM. Separately disposition stylesheet ::slotted()/other accepted pseudo-elements versus DOM queries before claiming evaluator support. Done when scope, specificity, featureless-host restrictions, fallback content and encapsulation precedence have explicit fixtures; DOM queries never pierce roots.
• SEL-03 - Language metadata fallback. Complete document/HTTP language metadata fallback behind the existing inherited lang/xml:lang matcher. Done when missing, empty and conflicting language sources follow the standard and results do not depend on process locale.
• HTML-01 - Parser custom-element construction remainder. Complete synchronous construction before attributes/children and constructor-failure semantics over the existing parser handoff and reaction queue. Preserve the completed insertion-reaction timing fixes. Done when constructors that observe attributes, throw, reenter or use document.write have native/ Browser/WPT evidence without adding another reaction channel or allowing two threads into the DOM.

Public API dependencies still required

• API-01 - Promote selectors. After SEL-01 and the accepted-selector capability audit, expose the immutable program/context/specificity and parse/match/query/closest operations in the selector contract. Done when an unsigned consumer proves namespace copying, errors, specificity, current-tree results, cancellation and cross-document reuse without accessing Browser/friend-only state.
• API-02 - Review public text declarations. Define a small text-based declaration contract before promotion; typed values and ParseCssValue are outside the renderless boundary. External mutations must preserve identity, bounded work and atomicity.
• API-03 - Promote mutable stylesheet/rule CSSOM. After API-02, review and expose the useful sheet/rule/list operations without exposing loader/page internals. Done when an unsigned consumer covers every advertised rule kind, ownership, insertion/deletion, descriptor mutation and serialization; existing public ParseCss remains a syntax-only API.
• API-04 - Review the standalone incremental HTML lifecycle. Design the smallest public contract over Html/TreeConstruction/HtmlParserSession: input ownership, append/advance/finish, observable partial trees, reentrancy, disposal, limits and cancellation. Done when these decisions and signatures are reviewed with executable internal contract cases; Browser script handoff and arbitrary callbacks must not become public by convenience.
• API-05 - Implement and publish incremental HTML. Depends on API-04. Preserve tokenizer/ tree-builder state across CRLF, surrogate, entity, tag/comment and raw-text chunk boundaries, with one operation-wide limit budget and well-defined EOF/failure behavior. Done when every boundary split matches one-shot parsing, cancellation/disposal cannot resume invalid state, and an unsigned consumer exercises the real lifecycle.
• API-06 - Review byte/stream APIs as a separate phase. Decide encoding sniff/restart, stream ownership, leave-open behavior, backpressure, partial-output and async cancellation before changing signatures or Browser transport. Done when the accepted contract and follow-on implementation tasks are explicit. If approved, implement them separately with multibyte/chunk/cancellation fixtures and proof that no script executes twice after a decode restart; string/incremental UTF-16 support does not close this gate.

Each public slice requires intentional API snapshots on net8.0/net10.0 and execution from the unsigned local-feed package consumer, without friend access or a project reference. A separate native pack/run is required for AOT evidence.

Acceptance still required

• Resolve the XML external-resource/no-fetch reviews and all eligible canonical-output comparisons. The complete 2,585-row census has 1,947 passing runnable cases, 27 verified optional policies, 593 outside-profile rows and 18 reviewed byte-boundary rows. All 386 eligible outputs are compared, including 66 reviewed no-fetch alternatives; unresolved cases, harness failures and known required-profile defects are zero. Missing-review, omission, notation, DTD PI and output corruption probes remain enforced.
• Finish the Swagger fixture. Its logo renders and the real summary click expands the operation through the Try-it-out control on both frameworks, with its original task and wait budgets.
• ACC-01 - Finish the real Scalar interaction. The pre-trim fixture reaches the GetEndpoint group and GET operation, then fails with System.TimeoutException: The operation's time budget elapsed. after the GET operation click, in Scalar's measure / onScrollChanged callback through getBoundingClientRect, PageLayout.ClientBoxOf, SizeQuery.HeightUpTo and selector matching. Next: isolate and bound the repeated geometry/cascade work with a deterministic regression, then rerun the interaction; do not remove the scroll handler or increase the task/wait budgets. Framework-caught failures require bounded console/request/DOM diagnostics; the console snapshot retains both its first and last messages so later timing logs cannot hide the initial exception. The fixture now also retains the first 20 warning/error console records separately from timing logs, and corrects two stale accessible button names without weakening their exact-match checks. An empty Page.Errors list alone does not prove success. Budgets remain unchanged. Rerun ScalarFixtureTests after each newly exposed fix. Done when the fixture opens the GetEndpoint group/GET operation, opens Test Request's API Client dialog, observes the expected OpenAPI request and reports no unexpected request/page/console failures on both frameworks.
• ACC-02 - Close named native Browser WPT debt. The document.write custom-element connection timing case is fixed. Two newly passing ordinary-parser reaction exclusions were removed; passing rows cannot remain excluded. Assign each remaining NeedsTriage subtest/cause to a concrete fix and regression; retain justified capability categories without mislabeling them implemented. Done when no unnamed failure or harness error remains and every exclusion matches failing tests but no passing tests.
• ACC-03 - Refresh the canonical Windows census. After the relevant ACC-02 fixes, run the documented JINT_WPT_BROWSER_CENSUS=1 check on Windows, review drift, then use =update only for justified reductions/equality refreshes. Commit both census/cause tables and refresh the inventory. Done when file/variant/registration totals and named causes reconcile, without raising failure ceilings or substituting macOS/Linux targeted results for the prescribed measurement.
• ACC-04 - Implement equivalent CSSOM comparison workloads. Extend the benchmark-only comparison infrastructure beyond HtmlParserComparisonBenchmark's complete-tree oracle. Separate syntax parsing, text sheet/declaration parsing, mutation and CSSOM reads; compare only the retained semantics, not full computed CSS against raw text. Done when untimed structural/semantic comparison and deliberate corruption probes reject mismatches in values, priority, parentage, ranges and mutations before any timing.
• ACC-05 - Complete paired performance acceptance. Depends on ACC-04 and the representative feature gates. Agree workloads/tolerances before measurement, then run the repository's default configured jobs in gate mode with alternating paired arms and control rows. Use at least six rounds for small effects; report confidence intervals, allocations and separate retained-memory evidence. Done when no unresolved supported regression remains, or a concrete tradeoff is explicitly accepted. No --job short numbers or invented speedup target.
• ACC-06 - External package and API acceptance. At every public promotion and the final checkpoint, verify both API snapshots, pack to a fresh local feed, run unsigned net8.0/net10.0 consumers and publish/run the Native AOT consumer. Until this item explicitly closes, Jint.HtmlParser remains a prerelease package independently of Jint GA: existing prerelease versions are retained; stable inputs acquire experimental-<BuildNumber> (default 0). All current public APIs, including positional record layouts, remain provisional and may evolve between prereleases. Browser's internal use and snapshot approval are not external acceptance. Stable promotion requires a reviewed removal of the project version policy with acceptance evidence. Done when package assets/dependencies/licenses are correct and each advertised API is exercised without signing, friend access, a project reference or a stale same-version package cache.
• ACC-07 - Full integration and dependency closure. After the feature gates, run complete native/Browser and relevant client/protocol fixtures on the supported frameworks, preserving Swagger, Scalar, XML comparisons and demand-parsing cache checks. Run dependency/binding staleness and inventory checks; keep AngleSharp only in comparison benchmarks. Done when remaining failures/skips have exact reviewed dispositions and the final evidence records commit, platform, framework, command and outcome rather than only aggregate pass counts.
• ACC-08 - Close the overall tracker. Reconcile the dependency inventory with the intentional text boundary and public API gate; link implementation evidence and accepted separate-phase decisions. Done only when all required items above are closed and ACC-01 through ACC-07 have current evidence. Targeted passing slices, dependency removal or this expanded plan cannot close the overall gate.

Routine checks are Release builds/tests, ParserDependencyTests, DomBindingsStalenessTests, the public API snapshots, the packed consumer and the dependency inventory. Raw CSS values are intentional; do not misrepresent them as full computed-style conformance.