Adopt shared controls and semantic styles across Buzz - #88

Post-#87 squash integration repair (2026-09-22)

Current head 6915669958ba5f0d6f5be1ec19bb9a533b2d3bf5 now targets main 09565c31ef3e125ac877025911028eae405b8172. #87 was squash-merged, so GitHub retargeted this PR and lost the shared ancestry. Main's tree is identical to the already-incorporated #87 parent 3b276793123aea43c23767d8c93f4e262b2a5872 (tree ea4480c473afd6bb6179da6502be13932b004f44).

This repair is a normal merge, not a history rewrite. It resolves five conflicts while preserving the prior #88 head a5669f2406f9f9820c1bf23d153b38ad3e8ee511 byte-for-byte (tree 401c2d153688cfc10b46baf0e8807ebe36959ba9). The new diff against main equals the previous diff against #87 exactly. No production, test, dependency or guard changes; all previous fixes remain.

Independent bounded source review confirmed both tree identities. Normal hooks at the new head passed app/viewer TypeScript, 2,017 unit tests in 193 files (17.90s), and all design/adoption guards. All 62 commits in the current PR range have sign-off trailers. No browser scenarios were added or removed; no browser/native rerun was added for this content-identical merge.

The prior head's hosted CI passed (run 35683088011); Windows native validation was skipped. That is evidence on the identical prior tree, not a fresh run on this merge commit. Fresh hosted CI and the existing formal requested-changes review remain separate gates. This repair grants no approval and does not establish native or comprehensive rendered visual/accessibility acceptance.

Earlier main refresh for morning review (2026-09-22)

Earlier head a5669f2406f9f9820c1bf23d153b38ad3e8ee511 retained the stack on refreshed #87 3b276793123aea43c23767d8c93f4e262b2a5872; both contain main 0beb523443557c189e5fcf16a3ac9e8437666690. Merge 890af02 resolves the four actual conflicts without rewriting history. Main's mention provider/wake notice, local-agent filtering, enrollment/cancellation and managed-agent page remain; #88's shared controls, admission explanations, sidebar memoization, shortcut-dismissal and thread-input-readiness fixes remain. The single AgentLibrary color edit follows main's extraction into its new owner.

The normal push hook exposed seven raw palette references in main's new agent editor. Follow-up a5669f2 maps them to existing semantic roles, seven substitutions across four files, with no dependency/behavior/guard/test changes. Error and warning colors retain their exact token values; agent-field focus changes from purple to the documented neutral border-focus role.

Validation:

• Merge tree 84391eb5d42e5a6f439769ff168757cd3c2dac65 passed 69 design tests, app/viewer production builds, 110 host cases across nine complete browser files (1.4m), and 38 viewer cases (21.4s), in Chromium and WebKit. Host files: mentions, typeahead, shortcuts, agent-control, plugin-import, design-system, appearance, agents, navigation-thread-history. The temporary viewer config changed only its occupied preview port and was removed.
• Final repair tree 401c2d153688cfc10b46baf0e8807ebe36959ba9 passed 32 cases in the complete agent-control, agent-models and agent-editor-grid files in both engines (37.2s). Normal final push hooks passed app/viewer TypeScript, all 2,017 unit tests in 193 files (17.28s), and every design/adoption guard. Committed repair exactly matches the tested diff. Every one of the 20 commits against refreshed Compose shared dialogs, tooltips and navigation #87 has a DCO trailer.
• Independent bounded source review cleared both conflict resolutions and the seven-token repair. No browser scenarios added or removed. Broader host/viewer evidence above predates only that seven-token repair; it was not repeated wholesale.

Both actual branch heads are pushed. This is an integration refresh, not approval, merge, native acceptance or release certification. Fresh hosted CI and existing formal requested-change review gates remain separate; historical evidence below belongs to its stated snapshots.

What changes

Buzz's shared components and semantic tokens now own common presentation across Settings, channels, Sessions, message actions, activity, pickers, workflows and media controls. Changing the design system updates those screens together. Base UI remains responsible for the relevant keyboard, focus and choice behavior; fonts, assets and fixtures remain suitable for the public repository.

This is 4 of 4: #73 → #85 → #86 → #87 → #88. Originally stacked on #87; now review against main after #87 merged. Current main's Sessions, drafts, unread state, agent admission, Phosphor icons, agent squircles, presence, Dock badge and development notification pause are preserved.

Adoption and review fixes

• Shared controls replace duplicated feature presentation. The new Dock permission actions use Button while retaining their existing request, pending, error and retry behavior.
• AgentChoice retains its avatar, name and caret through Base UI's render composition. Channel actions retain hover/focus visibility, unread emphasis, activity-dot size and destination attributes.
• Avatar owns its production mask import. A redundant feature import had generated a URL that returned HTML instead of the mask, hiding agent avatars. Real-browser decode checks cover both themes and engines.
• Tooltip text uses the paired semantic role. Button-title hints close when their trigger expands and allow movement to neighboring controls; explicitly descriptive tooltips remain hoverable. Download actions retain link semantics.
• Media stages pair surface-inverse with text-inverse. The role is limited to existing media stages, with contrast coverage.
• An adoption guard rejects production palette-ramp reads and feature CSS that repaints shared controls. Production CSS tests verify that token changes reach controls, chips and popovers. Its isolated hook fixture now supplies PostCSS and the examples directory, without bypassing any guard.
• Sidebar rows keep their connected controls inside a normal memo boundary. Unchanged rows no longer rebuild Base UI controls when another channel is selected. Real callbacks, session ownership, profile labels, working state, selection and unread updates remain live; no custom comparator ignores changed props.
• Shortcut tests wait for dialog removal and restored focus before the next shortcut. Typeahead tests observe accepted results/publication. Thread-opening tests wait for the virtualized list to accept pointer input before the existing real click. Pagination, reading-position, focus and modal-blocking assertions are retained.

The thread-input fix addresses a captured Linux WebKit failure: geometry had settled while Virtua still applied pointer-events: none. Playwright's alternate scroll alignment restarted the pointer lock, so the thread never opened. The scoped computed-style readiness check adds no delay, retry, forced click or product change. Independent review confirmed this lifecycle boundary.

Ownership that stays with features

The rich message editor owns caret, IME and completion behavior. Emoji Mart keeps its shadow-root adapter. Media renderers own media pixels, tile geometry, native zoom range, modal focus, dragging, playback and timecodes. Features retain layout, scrolling, data loading and subscriptions. Plugin compatibility aliases forward to shared semantic roles. This is shared presentation ownership, not a replacement for those feature contracts.

Validation

Historical head: 622990a1dce41613f7dcad9f565883596bd69b15, based on #87 1b5a4fcf597b9a44fd067e1ed1a3c3cde83aada8.

• Normal final pre-push hooks: 1,893 unit tests in 184 files pass, 16.50s wall / 89.30s summed execution, plus app/viewer TypeScript and all design/adoption guards. Formatting and diff checks pass; every PR commit has a sign-off.
• At the final head, the complete affected thread-history file passes 6/6 cases in Chromium/WebKit, 6.3s wall / 9.2s summed execution. The captured hosted trace provides failure evidence; macOS diagnostic samples did not reproduce the Linux pointer-lock window.
• The unchanged production tree at 7d01c09 passed static checks, production build, 17 mounted notification/Dock/sidebar tests, and 4 isolated opening cases across both engines, including main's 300-author thread. Warm switches were 38.5–40.8ms Chromium / 42–45ms WebKit against the unchanged 100ms limit.
• Hosted measurements at that same production tree passed 7/7 cases: 147.124s wall / 136.819s summed execution. Warm switches were 51.6–77.5ms, clearing the earlier 102.1ms failure. Slowest cases: cursor paging 65.858s, WebKit 300-author thread 20.778s, WebKit scrolling 17.075s. Local and hosted timings are separate environments, not a direct speed comparison.
• Earlier focused adoption checks passed 90 cases across eight full browser files in both engines; the sidebar performance fix passed 66 cases across ten full files (1.7min wall / 173.7s summed), plus six repeated isolated opening cases. These predate the latest main merge. The shared viewer passed 38 cases, and the repaired hook fixture passed 33 integration tests.
• All 12 hosted checks pass at 622990a: all four Chromium/WebKit journey shards, browser measurements, JavaScript, Rust/tool integration, Windows notifications, security, DCO and CI required. The previously failing WebKit shard now passes with the readiness fix. Final CI run. Native Dock/banner OS acceptance and human desktop visual acceptance remain separate from automated checks.

The sidebar improvement was measured using the same production-build fixture: four switches took 16.53/16.49/16.18/15.77ms of scripting before, versus 11.41/11.69/10.87/10.71ms after; parent #87 measured 12.36/10.66/10.51/11.23ms. This diagnostic profiling supports the ownership fix; it does not replace the unchanged acceptance clock or reduce the 128-row/1,001-profile fixture.

Browser coverage

This PR adds two engine cases, one shared-token production CSS journey per engine. Real stylesheet layers, inheritance and feature CSS interactions need a browser. Main's newly inherited presence scenarios remain unchanged. No existing browser journey is removed.

GIF retry/recovery, insertion, panel switching and query preservation remain covered. Avatar decode checks were added to existing light/dark activity cases. Existing thread activity catches the zero-size badge regression. Mounted regressions reproduced AgentChoice and hint failures before their fixes; sidebar tests cover callback replacement and session ownership. No browser budgets, engines, fixture scale, retries or correctness assertions were weakened to obtain passing results.