[Astro-Han]

Summary

• Add a browser-empty visual-smoke fixture scenario that mounts BrowserPanel (via a new VisualSmokeState.liveBrowserSessionIds field wired into applyVisualSmokeFixture) so the renderer chrome can be screenshot-captured in the empty state — the fix(ui): give BrowserPanel a narrow bottom-sheet layout (#546) #818 narrow-layout defect surface.
• Add a browser-panel-chrome source contract (symmetric to artifact-pane-layout.test.ts) locking the 5 declarative chrome invariants that don't need a screenshot: back/forward :disabled track canGoBack/canGoForward; reload/stop icon + aria-label + onClick swap with loading; address bar snaps to live URL on blur when not editing; empty state present when !hasPage.

Why

BrowserPanel (apps/desktop/src/renderer/browser-panel.tsx) is the renderer chrome over a native Electron WebContentsView; the native page can't be captured in a deterministic fixture, but the chrome is renderer DOM. The chrome had no deterministic visual or DOM/computed-style coverage, so layout/a11y regressions (e.g. the narrow-layout defect fixed in #818) could only be caught by manual real-window smoke.

Closes #819.

Scope

• Loaded / loading / nav chrome states are intentionally not screenshot-fixture'd. Their layout is identical to empty (fix(ui): give BrowserPanel a narrow bottom-sheet layout (#546) #818 is a layout defect, not a state defect), loading/nav are transient and can't be deterministically captured, and their state invariants are locked by the source contract. Seeding BrowserState for a loaded screenshot would need a production test-seam (a main-side visual-smoke browser:state emit hook, or a visual-smoke read inside BrowserPanel) for no layout-regression value — rejected on first-principles/Occam's-razor grounds.
• Source contract over render test. The repo has no jsdom/testing-library, and BrowserPanel is IPC-driven via window.maka.browser. The 5 invariants are declarative (disabled={expr}, cond ? <X/> : <Y/>, onBlur={...}, !hasPage && <Empty>) — the wiring IS the rendered DOM, so a source regex is a reliable proxy without a render harness + heavy mocking. This matches the cited symmetry (artifact-pane-layout.test.ts).
• liveBrowserSessionIds is a natural seam. It's the exact state the renderer already uses to gate BrowserPanel mounting (activeId && liveBrowserSessionIds.includes(activeId)); adding it to VisualSmokeState extends the existing fixture contract, no production behavior change. Real users never receive a visual smoke state.

Verification

• typecheck ✓ (main + renderer + storybook)
• test ✓ — 2363/2363 (was 2354; +2 browser-empty fixture tests, +5 browser-panel-chrome contract tests)
• browser-panel-chrome ✓ 5/5 (back/forward disabled, reload/stop swap, address blur-snap, empty-state)
• visual-smoke-fixture ✓ — browser-empty scenario resolves, activeSessionId + liveBrowserSessionIds seeded correctly, turn session on disk (reused, no browser-specific seed)
• 8 baseline PNGs generated at apps/desktop/tests/screenshots/browser-empty/ (gitignored generated captures): 1280/990 × light/dark × motion/reduced. Dimensions verified via sips (1280→2560×1640, 990→1980×1640 Retina; sizes 249KB light / 729KB-1MB dark, non-blank). motion == reduced-motion byte-for-byte (empty chrome has no animation).
• Visual evidence: I can't read images, so I can't visually verify the PNGs. The chrome layout regression surface is locked by the existing browser-panel-layout CSS contract (fix(ui): give BrowserPanel a narrow bottom-sheet layout (#546) #818/fix(ui): reserve chat/composer height when both ArtifactPane and BrowserPanel are open at narrow widths #824, already merged) + the empty-state fixture; the 5 behavioral invariants are locked by the source contract. The generated PNGs are a stable-subset candidate pending visual review + promotion to screenshots-baseline/ via a separate baseline commit (the established pattern — baselines are committed separately from fixture code, e.g. e1d1344e, 12ad19aa; --update-baseline is designed for a full --all re-capture, not single-scenario promotion).

Impact

None on users — visual-smoke mode only; real users never receive a visual smoke state, so liveBrowserSessionIds never drives production browser:live wiring. Dev/test impact: a new fixture scenario + a new source contract test; the capture script's ALL_SCENARIOS includes browser-empty for baseline runs.

Reviewer notes

• The browser-empty scenario reuses the always-seeded turn session (TURN_SESSION_ID), so it adds no on-disk seed branch — getVisualSmokeState returns liveBrowserSessionIds: [TURN_SESSION_ID] + activeSessionId: TURN_SESSION_ID; seedVisualSmokeFixture already writes the turn session unconditionally.
• Per-capture wall time was ~60s (right at the capture script's 60s subprocess timeout) — the app doesn't self-exit after the capture marker, so the script kills it at the timeout; all 8 variants still produced valid PNGs. This is pre-existing capture-script behavior, not introduced here.

Ready for review

• Verification is complete and the results above are current
• Impact, compatibility, migration, documentation, and release-note needs are addressed
• User-visible UI/UX changes include visual evidence, or Verification explains why it is not applicable