Complete cascade layers, origins, global keywords, and custom-property resolution

[wieslawsoltes]

Parent tracker: #235

Problem

CSS correctness requires the same winning declaration and computed value as a browser across origins, importance, layers, specificity, inheritance, shorthand expansion, global keywords, and custom-property dependencies. Current contracts cover important paths but not a complete interaction matrix.

Proposed work

• Add a table-driven cascade oracle fixture compared with Chromium.
• Cover author/user-agent origins available to WebScene, !important, source order, layers, specificity, inline declarations, inheritance, and shorthand/longhand interaction.
• Complete initial, inherit, unset, and revert behavior at specified/computed-value boundaries.
• Cover custom-property fallback, cycles, invalid-at-computed-value-time behavior, and dependency invalidation.
• Cache/compile only after equivalent results and invalidation lifetimes are proven.

Acceptance

• Browser/native winner and computed-value matrices match.
• Dynamic declaration, class, ancestor, and custom-property mutations recascade correctly.
• Cycles and fallbacks fail deterministically without stale values.
• Bounded cascade counters and timing evidence accompany any optimization.
• Fix catalog script and logical sizing compatibility #148 behavior is treated as the baseline and [Epic] CSS features, compatibility, validation, and performance #235 is updated.

[danwalmsley]

Coordination from #245: a production Spotify breakpoint profile found repeated pseudo-element/host classification in the per-element candidate loop. I am moving only that immutable classification into the existing interned rule payload and reusing its already-compiled pseudo origin. No cascade-layer/origin/global-keyword/custom-property algorithm changes. Shared native service and Chrome/native stylesheet-replacement + shadow-host/generated-box regressions pass. Timing comparison is pending; broader #238 acceptance remains open. Separate gap found while writing the test: getComputedStyle(element, pseudo) ignores its second argument; replacement tests use generated layout geometry instead and do not claim that API is supported.

[danwalmsley]

PR #245 integration follow-up, pushed 86c83845: the packaged-SDK Kestrel parsed/compiled gate found a regression in the earlier immutable rule-classification change (51fef1f2). Runtime parsing populated pseudo_kind / host_selector, but webscene-uic's prepared-sheet emitter omitted both fields. Generated pseudo/host rules consequently used zero/default classification. The fix emits both immutable fields; it does not alter cascade precedence or matching semantics.

The old installed compiler fails the new controlled emission regression. The fixed compiler suite and parsed/persisted/precompiled shared-stylesheet metadata + scene equivalence tests pass. Generated-box CSSOM replacement coverage passes Chrome/native 6/6. Fresh clean SDK and source-denied Kestrel qualification are being repeated, so the prior merged SDK is not yet accepted for parsed/compiled parity.

While constructing the browser oracle, a separate unitless-line-height discrepancy was found: an empty flex origin with font-size:10px;line-height:1 and :before {content:'x';font-size:40px;line-height:1} should have height 40; Chrome passes and native fails. It is not caused by the compiler-emission omission (the runtime-only contract also fails). Retained diagnostic: /tmp/css-compiled-pseudo-{chrome,native}-20260917/results.json. The compiler regression uses explicit generated-box geometry instead; no workaround for font metrics was introduced. Please treat this as a separate generated-content/computed-value follow-up in #238/#240, not a second implementation started in those lanes.

Current production sampling still attributes breakpoint work to compound matching and candidate sorting. This run is not an accepted timing comparison: live content changed to 2,726 elements / 18 images and other host compilation was active. No Chrome-speed or physical-presentation claim; no CI monitoring.

[wieslawsoltes]

Reopening from exact merged-main package evidence. PR #300's focused gates passed, but the broader required RID matrix is red in post-merge run 35252631301, and the same failures remain on current main 70a2db4c in run 35254142552.

The #238 acceptance concern is the Linux CSS runtime evidence: webscene_css_containment_tests misses its 1 s completion bound; the 4,096-control/100-cycle accessibility gate takes 20.53 s against 15 s; and CSS effects misses its bounded wait. These failures reproduce before #298. A before/after bisect is still required before assigning every timeout to the layered-cascade implementation, but #238's bounded cascade timing acceptance is not proven while the required package matrix is red.

Two other failures in those jobs are tracked separately and must not be attributed to PR #300 without a bisect:

• macOS floating-point std::from_chars in webscene_native_dom_scene.inc predates PR Implement layered CSS cascade winner matrix #300 (introduced by retained inset-clip work) and needs its own portability repair.
• Windows webscene_indexeddb_compatibility.h raw-literal C2026 predates PR Add readable response and FetchEvent vertical #298 and is tracked by Keep the IndexedDB bootstrap below the MSVC literal limit #301.

The full Linux suite also reports the identical-resource four-engine invariant and IndexedDB same-window-navigation rejection. Attribution must separate cascade-induced starvation/state leakage from independent baseline defects.

The existing #297 stylesheet-finalization candidate is being rebased independently and must not hide synchronous CSSOM or cascade-layer semantics. This issue closes again only after the original 17/17 cascade oracle, containment/accessibility/effects budgets, and the required Linux package job pass on the same focused head. Cross-RID portability and unrelated resource/IndexedDB defects retain their own owners.

[wieslawsoltes]

Taking the bounded Linux timing-qualification slice from exact main 348ea440501c4059da6f95d1fbd6f6159a5df47b.

Live overlap audit: the only open PR is docs-only #76. This slice does not change #257 live form state, #338/#339 media dependency metadata, stylesheet ownership/indexing, cascade winner ordering, global-keyword/custom-property semantics, or generated-content/layout behavior.

Before-edit attribution from the package runner:

• preceding main bea90e8e (run 35288000024): containment 1.07 s generic script timeout, accessibility 17.32 s against the 15 s gate, effects 15.71 s generic script timeout;
• current-main-equivalent tree 0d6d95a8 / merge tree 348ea440 (run 35318401587): 1.08 s, 19.11 s, and 15.72 s respectively.

Containment and effects have dispatch waits shorter than their explicit measured performance budgets (1 s vs 10 s and 15 s vs 25 s). Their focused workloads complete inside the real budgets once allowed to report measured results. Accessibility is causal hot declaration replay: the 4,096-control/200-state trace spends about 2 ms planning and about 30 ms applying each state on an otherwise idle local run; direct sampling places the application CPU in supported-property declaration handling.

Owned candidate boundary:

• preserve exact certification feature-use observations while caching the repeated supported CSS-property writer lookup without rebuilding five strings plus a composite key for every declaration;
• align the two dispatch waits with their existing explicit performance assertions, retaining the stricter measured gates;
• retain numeric elapsed time, DOM lifecycle, V8 heap, textual-style storage, scene size/balance, and retained-node evidence.

Before merge I will require the unchanged 17/17 cascade winner matrix in current and legacy parsers, direct Chromium 17/17, dynamic/custom-property/media regressions including #257/#338/#339, the three focused Linux gates on the same exact head, and exact feature-use count/classification coverage. Authoritative performance measurements wait for the concurrent Release build to finish so host contention is not mistaken for a regression or gain.

[wieslawsoltes]

Accepted and merged via #346 at 114f38e1b5e5f9a156a2a8aadca2288d94e59db1.

Exact PR head 2fdac5a6 passed the directly related Linux package gates in run 35323977485, job 105532770042:

• containment: 2.61s, pass;
• accessibility: 13.62s, pass against the 15s 4,096-control/100-cycle bound;
• effects: 26.82s CTest wall time, pass; its distinct <25s scale and <15s publication assertions both passed.

The final current-parser and legacy-parser cascade matrices remain 17/17, matching the recorded Chrome 152 oracle. Dynamic form state, layer scaling, media/custom-property/inheritance invalidation, feature-use counts, host-clock transitions, heap/RSS, and lifecycle checks pass. The amended local accessibility median is 5.088s, 33.6% below the prior candidate's 7.661s median.

The aggregate Linux job's remaining failures are the separately owned identical-resource four-engine invariant, IndexedDB navigation, and ServiceWorker parser timeout. They do not touch or invalidate the three #238 acceptance gates.