test(desktop): add rendered computed-style contract for the chat chrome

[Astro-Han]

Summary

Closes#1312.

#1308 deleted the chat-chrome pixel script; the surviving chat-chrome-no-gradient-contract.test.ts regex-matches the first source rule per selector, so it cannot see the rendered cascade — a real blind spot: the "three-color-band" incident documented in theme-glass.css was exactly a checked declaration losing to an unlayered higher-specificity rule in reference-shell.css.

This locks the rendered invariant with a Playwright e2e spec, apps/desktop/e2e/chat-chrome-style.spec.ts, measuring the REAL shell elements of the live app (fake backend, long-transcript fixture). The platform axis boots natively: the existing e2e-fixture override seam (which already carries theme/locale/timezone/reduced-motion) gains a MAKA_E2E_FIXTURE_PLATFORM override that app:info reports to the renderer, so a fixture window enters the darwin glass cascade — or the opaque win32 base cascade — through the production data-os path on any host. The theme axis drives page.emulateMedia({ colorScheme }) — the real prefers-color-scheme path the fixture's seeded theme: 'auto' production listener follows — because an in-page .dark class flip leaves the relative-color chains in --surface-canvas resolved against stale token values (measured: appFrame stuck at light [247,247,247] instead of dark [9,9,11]). Reads wait two animation frames after the class lands (the relative-color recalc settles only after a produced frame), and a per-window guard asserts the light and dark appFrame reads differ — equal reads mean a phantom stale cascade, so an insufficient settle fails loud instead of passing vacuously.

Per combo, in the expanded sidebar state, the spec asserts: the effective gutter (Porter-Duff composite of .appFrame ← .maka-shell-2col, the only two painters of the gutter pixel) is fully painted (alpha 255) and visibly distinct from the effective card; no background-image on any chrome surface (the thrice-rejected 172deg gradient is invisible to color reads); no seam on either boundary side including both resize-handle borders; transparent handle background and resting ::after grip; no adjacent box-shadows; exact-string '12px' corner radii (catches elliptical 12px / 0px; doubles as the vacuous-green self-check); exact-string '4px' card margins; a mode-split shadow lock — dark requires a layer with visible color alpha AND non-zero paint geometry, light requires exactly none per the reference-shell.css design comment; cumulative ancestor-chain opacity exactly 1 (an opacity: 0 anywhere hides the shell while every color read stays green); resting outlines invisible and filter: none on the seam trio (outline and drop-shadow() paint boundary lines the border/box-shadow checks cannot see).

Shape history, for reviewers: four external review rounds (Codex + Kimi K3, fresh sessions each) drove this from a standalone synthetic-DOM Electron script through an in-place-clone probe to the current form. The through-line: any probe state — hand-built mirror DOM, shallow clones, or a cascade reached by an in-page class flip — diverges from what a real user's window computes, in ways that hide regressions (missing painters, relational-selector blindness, stale relative-color chains). Round 4 proved the previous shape's dark combos asserted against a phantom shell (the historical dark transparent-backplate regression passed silently) and replaced the flip with the production media-query path; it also removed a guard that bound a visually-equivalent internal painter split (darwin's residual transparent shell) instead of final appearance. The source-string contract stays as the fast pre-check; its header points here.

Verification

• Spec green: both forced-platform windows × both emulated themes in one run (2.9s); full e2e suite 33 passed; desktop unit tests 2764 pass (including the platform-override cases: closed enum, fail-closed to null, dev/test-only); typecheck clean.
• Non-vacuity — five red proofs this round (source edit + rebuild + rerun + exact restore, never committed), each targeting a vector the prior shape provably missed:
  • html.dark .appFrame, html.dark .maka-shell-2col { background: transparent } → dark gutter alpha fails RED (this exact regression stayed GREEN against the in-page-flip shape — the decisive phantom proof);
  • deleting the settle → the per-window phantom guard fails RED (light=[247,247,247] dark=[247,247,247]);
  • .maka-shell-2col { opacity: 0 } → ancestor-opacity fact fails;
  • resting outline: 2px solid red on the handle → outline fact fails;
  • filter: drop-shadow(0 0 4px red) on the panel list → filter fact fails.
  • Earlier rounds' red proofs (172deg gradient, shell recolor, elliptical radius, margin: 0.01px, zero-geometry dark shadow, late 1px border, adjacent-sibling border, ::after grip color) remain covered by the carried-over assertions.
• Production-code footprint: the platform override rides the existing e2e-fixture override seam — one fallback in the app:info handler (e2eFixture?.platform ?? process.platform), one parse function with the same closed-enum fail-closed discipline as theme/locale/timezone, active only in unpackaged fixture mode.

Review focus

The spec header documents the gutter painter model, why the platform axis must be booted (no mid-session data-os value flip on a long-lived resolved cascade) and the theme axis must ride the real media-query path (an in-page class flip reads a phantom shell — measured values in the header), and the settle + phantom-guard pair that keeps the dark reads honest on any host.