Close remaining CSS layout gaps with browser-referenced geometry contracts

[wieslawsoltes]

Current layout checkpoint — 20 September 2026

Exact merged heads: WebScene a46b62e14eb197603f7e227c1558afb9088633b0; AppScene 97702daeb9ad71341ced7226f0aa966dc77ea53c; unchanged Code OSS 645f29cc3176500b4b5762ba887cf2a7f0ffdf2c; local vscode-demo e4b98e5470cd7856aff3dfd6843c8e125d7db933 (clean, intentionally unpushed). Open consolidations remain WebScene #76 085bba760fa571c8f08ea866dc5dd1ec27c4d376, AppScene #65 ce66ef6ceb776b3f14e129f853bfc08b5c5e608c, and vscode-demo #1; do not merge them. One-agent mode is active.

Focused layout slices recorded here are merged through the later isolation, text-wrap, word-break/overflow-wrap, retained image fit/alignment, overscroll and touch-action work now present on current main. No focused layout branch is active.

Continue implementation only from a reduced current-main geometry defect. Browser/native geometry, mutation scope, 4,096-node bounds, editor/Settings/Explorer/terminal pixels, package and cross-RID acceptance remain deferred.

Parent tracker: #235

Problem

Remaining CSS layout gaps surface as product-specific visual differences. We need browser-referenced, product-neutral geometry contracts across the layout algorithms used by desktop applications.

Proposed work

• Inventory layout features used by unchanged VS Code after Fix catalog script and logical sizing compatibility #148 lands.
• Expand contracts for physical/logical axes, intrinsic/min/max sizing, percentage resolution, box sizing, margin/padding/border, absolute/fixed/sticky positioning, overflow/scrolling, flex, grid, tables, and native form controls.
• Compare bounding rectangles, scroll metrics, baselines, and computed values with Chromium.
• Exercise dynamic insertion/removal, resize, writing-mode changes, and inherited/custom-property inputs.
• Add layout/scene counters and representative performance budgets.

Acceptance

• Each retained fix has a minimal product-neutral browser/native geometry regression.
• Dynamic mutations and viewport changes recompute the correct scope.
• VS Code editor, Settings (Match absolute modal and nested Settings layout used by VS Code OSS #112), Explorer, terminal, and panel geometries pass their scenario contracts.
• No application-specific CSS workaround is added.
• [Epic] CSS features, compatibility, validation, and performance #235 maintains the ordered list of remaining algorithms.

#148 owns its current logical sizing, responsive grid, scrolling, and catalog layout changes; do not duplicate them.

[wieslawsoltes]

Focused implementation candidate is ready on current main d7720a63: commit 01c12edac7e4186b1eb1930d5b4270e698f7cd63, branch fix/generated-content-line-height-240.

The reduced Chromium/native contract isolated the generated-content geometry gap. For a 40px ::before, native stored unitless line-height:1 and line-height:5 as literal pixel lengths, producing heights [60,1,5,2]; Chromium produces [60,40,200,2] for inherited 1.5, authored 1, authored 5, and explicit 2px. The fix preserves <number> as a multiplier until the generated box font size is known, uses the same representation in retained text publication, and removes the old paint-only “positive values <=4 are multipliers” heuristic. Dynamic class mutation restores 40→60→200→40 geometry and retained scene strings carry exact 40/60/200/2 line heights.

Exact-current gates:

• Chromium authority: 1/1 documents, 5/5 subtests, 324 ms.
• Native WPT: 1/1 documents, 5/5 subtests, 282 ms.
• Focused native dimension-variable-compatibility: pass, including adjacent font-relative/layout cases and retained text-resource assertions.
• 4,096 generated boxes, 20 mutation/layout cycles: p95 115.67 ms, max 130.74 ms; V8 heap 678,776→678,776 bytes after low memory; peak RSS 40,845,312→61,784,064 bytes (20.0 MiB growth, 64 MiB bound); DOM node and pseudo storage return to baseline bounds.

The four changed paths have zero overlap with merged #312’s parser/resource paths and do not touch #286’s CSS parsing/cascade or oracle/profile paths. This is a focused #240 tranche; it does not close the broader layout issue.

[wieslawsoltes]

The generated-content unitless line-height slice is complete and merged via #322 as 4ccadd1333c54e91e83e65234e783685964ae1d9.

This closes the measured gap where a 40px pseudo-element incorrectly treated unitless values as pixels: Chromium/native now agree on [60, 40, 200, 2] for inherited 1.5, own 1, own 5, and explicit 2px. Dynamic restoration, retained text-resource geometry, browser/native WPT, the V8-free shared CSS service, a 4,096-node × 20-cycle bound, and the exact Native Linux document/compiler gate pass. The final profile was p95 106.36 ms, max 106.70 ms, unchanged V8 heap, and +19.98 MiB RSS.

#240 remains open for its other independent layout gaps.

[wieslawsoltes]

Implementation merged via #444 at d28995e2a8cd898a6ba5a3f43baffe091ccda18a from exact source 597b387f69a2346808faea308e6ef6c9e63bea02 on base ac235def2417569b949af14ff18fd99153c88dbd.

The focused slice exposes and retains aspect-ratio, carries it into generated pseudo boxes, applies a preferred ratio when one axis is definite in ordinary/flex/grid/floating/positioned layout, accounts for content-box versus border-box edges, serializes computed style, and includes ratio changes in layout equality.

No validation was run under the implementation-only instruction. Outstanding debt remains exactly the browser/native geometry matrix, invalid-value and min/max transfer coverage, dynamic mutation and pseudo coverage, the 4,096-box performance/heap/RSS/lifecycle/idle-publication gate, and exact-package Code OSS editor watermark, banner, and aquarium regions. This issue is reopened to retain that debt after the PR's Closes directive automatically closed it.

[wieslawsoltes]

Implementation-first re-audit at WebScene 488b9790 found no documented browser-reduced product-used layout defect beyond implemented owners. #439 aspect-ratio behavior is merged; its remaining geometry/mutation/performance/lifecycle/package matrix is validation debt. The developer-owned #112 worktree was left untouched. No code or runtime validation was run; git diff --check passed.