Establish Chromium/AppScene visual and geometry regression gates for unchanged Code OSS

[wieslawsoltes]

Current visual-gate checkpoint — 20 September 2026

Exact merged heads: WebScene a46b62e14eb197603f7e227c1558afb9088633b0; AppScene 97702daeb9ad71341ced7226f0aa966dc77ea53c; unchanged Code OSS 645f29cc3176500b4b5762ba887cf2a7f0ffdf2c; local vscode-demo 123ad63a345dd77ceb0d75aed4c8eda244cf1288 (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.

#799/#806 is merged for the 52 touch-action declarations across 27 unchanged Code OSS files. Exact visual/geometry qualification must cover contact scrolling and cancellation on tabs, editors, split views, sliders, terminal, and Chat while proving mouse rendering remains unchanged. No exact product visual, package, hardware-contact, accessibility, lifecycle, memory, or performance execution ran.

Parent epic: #227.

Problem

The integration has focused screenshot checks for individual fixes, but it does not yet have one revisioned Chromium/AppScene oracle covering every unchanged Code OSS surface. Without a shared harness, a feature can become functional while geometry, fonts, clipping, focus, or retained-scene output regresses elsewhere.

Investigation first

Before implementation:

1. inventory every existing screenshot, geometry, DOM-state, scene-counter, and pixel-budget probe in SceneTech/vscode-demo and WebScene;
2. record exact gaps for editor, Welcome, Explorer, Search, Source Control, Run/Debug, Extensions, Settings, Command Palette, Problems/Output/Terminal, Markdown/webviews, Chat, notifications, menus, dialogs, and multi-panel resize;
3. define stable upstream state, viewport, DPR, fonts, theme, locale, animation policy, profile, fixture data, and capture readiness for both Chromium and AppScene;
4. link any reproduced engine defect to its focused issue rather than fixing it in the harness.

Proposed direction

Build a data-driven acceptance manifest that drives the unchanged Code OSS server in Chromium and the same payload in AppScene. Compare semantic DOM state and named-region geometry before bounded perceptual/pixel comparison. Retain compact crops, measurements, hashes, and failure diffs; delete superseded full traces and packages.

Dependencies and boundaries

• Add native input observability and Code OSS activation gates #100 owns input/activation observability.
• Match absolute modal and nested Settings layout used by VS Code OSS #112 owns Settings/select layout.
• [Epic] CSS features, compatibility, validation, and performance #235 and its children own CSS/layout/paint fixes.
• Render generated icon-font pseudo-elements correctly in tabs and list rows #246 owns the active-tab/Explorer SVG artifact.
• Qualify remote workspace bootstrap, Explorer contents, and watcher refresh #252 owns workspace/Explorer content.
• Webview/Markdown tickets own nested-document rendering.
• This issue owns the common oracle, readiness rules, manifests, evidence format, and release-wide execution.

Acceptance gates

• Every listed surface has deterministic setup, readiness, semantic assertions, named geometry, and a justified pixel budget.
• Light/dark, 1x/2x DPR, resize, active/inactive focus, and at least one high-contrast/forced-colors state are represented where applicable.
• Missing fonts/resources, blank regions, stale frames, clipped controls, unexpected overlays, and wrong selected/focus state fail closed.
• Baselines name exact VS Code, AppScene, WebScene, SDK, host OS, font, and Chromium revisions.
• Harness overhead, retained artifacts, run time, and peak disk use have explicit budgets.
• A final unchanged Code OSS Release passes before promotion.

Proposed PR stack

1. inventory + manifest schema + deterministic Chromium oracle;
2. AppScene semantic/geometry capture + readiness checks;
3. bounded pixel/diff evidence + storage budgets;
4. full-surface release manifest.

Validate the cumulative top of each dependent group and merge focused WebScene fixes separately from the vscode-demo consolidation PR.

[wieslawsoltes]

CSS ownership/dependency audit against #235:

• This issue owns the shared release-wide manifest, readiness protocol, Chromium/AppScene semantic/geometry/pixel capture, and compact diff evidence.
• Profile unchanged VS Code CSS and visual-tree workloads with bounded performance gates #243 owns unchanged-product CSS/visual-tree profiling, while Bound visual-tree churn, idle publication, and memory across panel mutations #258 owns the reusable panel-mutation counters, idle-publication, retained-node, latency, and RSS-slope workload. Consume their measurements rather than duplicating those harnesses here.
• Qualify forced-colors, system colors, focus visuals, and motion preferences #255 owns forced-colors/focus/reduced-motion CSS behavior; Implement CSS containment, container queries, and content visibility for dynamic panels #254 owns containment/content-visibility behavior; Render CSS masks, filters, clip paths, and backdrop effects in retained scenes #256 owns masks/filters/clips/backdrop paint; Render generated icon-font pseudo-elements correctly in tabs and list rows #246 remains the focused tab/list SVG parity gate. Qualify Code OSS accessibility semantics, keyboard navigation, and native exposure #262 owns accessibility semantics/native exposure, and Publish a Code OSS Web API capability ledger and expand WPT coverage #263 owns the global capability ledger. Route a failed capture to those owners.

Missing cross-cutting gate to carry in the manifest: where a focused owner exposes retained-scene/resource checksums, store and compare those alongside DOM, geometry, focus/accessibility, and pixels. Report explicit pass/fail/skipped/unavailable/total denominators and exact source/browser/font/platform revisions; a missing required mark or baseline fails closed.

[wieslawsoltes]

Wave 0 manifest foundation committed locally

vscode-demo local commit c367693 adds the fail-closed release inventory required by this issue without changing VS Code OSS:

• config/visual-scenarios.json declares 16 deterministic states across editor, navigation, workspace, panels, webviews, Chat, and dialogs;
• every state names its owning AppScene/WebScene issues, capture state, acceptance state, theme/DPR/focus variants, and semantic/geometry/pixel/focus/accessibility/performance/lifecycle evidence;
• the manifest is pinned to the exact VS Code revision in config/upstream.lock.json and rejects stale provenance;
• scripts/visual_scenario_manifest.py rejects missing/duplicate evidence dimensions, silent skips, invalid issue ownership, missing or mismatched active contracts, undeclared group membership, invalid paths, and oversized reports;
• the machine report exposes explicit scenario denominators (1 passing / 4 partial / 6 blocked / 5 pending / 16 total) and evidence denominators (11 gate / 101 planned / 0 not-applicable / 112 total);
• existing editor, Command Palette, and Settings gates can now select their active contract with visual-parity.py --scenario ...; selecting a planned state such as Markdown fails instead of pretending a capture exists;
• per-run/report/retained-run artifact budgets are versioned in the manifest.

Validation on exact local inputs:

• Python repository suite: 141/141 passed;
• Node repository suite under required Node 24.18.1: 65/65 passed;
• focused manifest + visual tests: 22/22 passed;
• generated report: 3,539 bytes, inside its 1 MiB budget.

The system /usr/local/bin/node is Node 16 and predictably cannot run this repository's modern Node tests; the pinned Node 24.18.1 result is the valid gate. The commit remains local/unpushed under the consolidation policy. Next #259 work is implementing deterministic readiness/capture contracts for the planned scenarios rather than loosening existing baselines.

[wieslawsoltes]

Welcome scenario capture contract implemented locally

vscode-demo local commit ec6f0b2 advances the #259 manifest from three to four implemented capture scenarios without modifying the pinned VS Code OSS source.

The new code-oss-welcome-parity-v1 lane:

• waits for the real code/didLoadExtensions workbench mark before opening the upstream Welcome editor;
• requires three stable semantic samples with Code - OSS, Editing evolved, New File..., at least two start actions, at least one walkthrough, and the exact startup checkbox label;
• compares named geometry for the Welcome container, header, left/right columns, start list, walkthrough list, and footer;
• evaluates pixel drift only inside the measured Welcome container;
• binds the capture to the existing transient-profile, exact-payload, clean AppScene/WebScene provenance, artifact-hash, and disk-budget audit;
• marks Welcome capture as implemented while leaving acceptance pending until a fresh current-head Chromium/AppScene package run passes.

Current manifest denominators are 1 passing / 4 partial / 6 blocked / 5 pending / 16 total and 14 gate / 98 planned / 0 not-applicable / 112 total evidence obligations.

Validation:

• complete Python suite: 158/158 passed;
• complete Node suite on exact packaged Node 24.18.1: 66/66 passed;
• focused visual/manifest/packaging Python suite: 56/56 passed;
• Welcome-aware browser/overlay Node suite: 20/20 passed;
• overlay applied to pinned unchanged VS Code OSS 645f29cc, full TypeScript no-emit compile passed, and the overlay reverted to a clean submodule;
• git diff --check passed.

The commit remains local and unpushed under the consolidation policy. The next #259 step is a fresh package after the active upstream stacks merge, followed by retained Chromium/native Welcome evidence and the remaining focus/accessibility/performance/lifecycle obligations.

[wieslawsoltes]

Local vscode-demo commit 9721499 adds a deterministic Chromium/AppScene visual lane for the reported malformed document-tab background. It remains intentionally unpushed on integration/appscene-webscene-code-oss until the next consolidated update.

The query-gated unchanged-Code state now:

• closes the startup editor and creates three real dirty untitled Monaco editors;
• requires the exact ordered labels Untitled-1, Untitled-2, Untitled-3, with the third active, all three dirty, and three stable samples;
• records named geometry for the complete tab strip, each tab, the active tab, and surrounding workbench surfaces;
• restricts pixel metrics to the measured editor tab strip, so the path-like SVG/background artifact cannot disappear into a whole-window tolerance;
• adds a fail-closed code-oss-editor-tabs-parity-v1 contract and package marker.

The visual inventory now contains 17 scenarios and 119 evidence obligations; six captures are implemented. Focused and complete validation passes: Python 164/164, exact Node 24.18.1 68/68, manifest audit, and the applied pinned VS Code TypeScript no-emit build. The VS Code submodule was reverted byte-exactly and is clean.

Acceptance remains partial until a fresh current-head Chromium/native capture supplies semantic, geometry, and pixel evidence. Focus, accessibility, dynamic performance, and lifecycle remain owned by their existing #258/#260/#262 lanes.

[wieslawsoltes]

#246 changes the product acceptance target based on direct Chromium inspection. The visible active-tab and Explorer artwork is generated Seti icon-font content, not an SVG background. A diagnostic post-#251/#128 capture still shows both shapes while the old whole-window contract passes at MAE 0.005242, RMSE 0.048151, and changed-pixel ratio 0.010378.

Keep code-oss-editor-tabs-parity-v1 as the tab acceptance lane, but require the active tab icon slot as named geometry/pixel evidence. Add a separate deterministic Explorer selected-row/icon region: the current three-tab region ends at the tab strip and cannot qualify the cyan Explorer glyph. Run both only from a fresh exact current-main AppScene/WebScene package after the WebScene pseudo font-family fix and its product-neutral native/WPT gates pass. Full primitive, source-path, revision, and hash evidence is on #246.

[wieslawsoltes]

Issue #246's underlying generated-icon defect now has local WebScene candidate 6e6d5568. Exact main fails the dedicated PUA native reftest by 21,796 pixels while Chromium passes at zero; the candidate passes native and Chromium with exact per-renderer test/reference hashes. After focused review/integration, schedule one exact-package run for code-oss-editor-tabs-parity-v1 and a separate deterministic Explorer selected-row/icon region. Keep #246 open until both product regions pass; no AppScene or VS Code stylesheet change is needed.

[wieslawsoltes]

The generic fix for the fresh Explorer path-like artifact is merged through #418 at eb3cfe38f13dfe1243404ed87af39bf77eaf902a (child #417 of #246).

Root cause: the pre-fix WebScene package retained Code OSS's pseudo font-size:150% as 150px instead of resolving it from the 13px originating label to 19.5px. The Explorer row's 22px clip made the oversized icon glyph look like a row-spanning path. This was not a row background/clip defect.

The reduced browser oracle is pixel exact and the native scene gate now asserts 19.5px with the existing 4,096-icon mutation/performance/lifecycle coverage. Please use a WebScene build at/after eb3cfe38 for the next unchanged-Code-OSS Explorer and editor-tab package capture. The pinned pre-fix vscode-demo checkout was left unchanged.