refactor: shrink the renderer's first-load chunk and remove three duplicated authorities

[Astro-Han]

Five simplification-audit findings in the desktop renderer and packages/ui. All internal; no user-visible behavior changes.

This is not a fix for the session-switch jank. The measured session-switch cost (9 long tasks, 717 ms, CLS 0.55) has a different, still-unidentified cause and is tracked separately. What is here is bundle size, one duplicated authority removed, and two boundary escape hatches closed.

Bundle size

The workbar feature barrel statically re-exported WorkbarSurface while workbar-host reaches it through lazy(() => import('./workbar-surface')). Because app-shell imports WorkbarHost from that barrel, the static edge dragged the 917-line surface and its five nested tool panels into the eager chunk — rolldown reported INEFFECTIVE_DYNAMIC_IMPORT.

Two Storybook files did import WorkbarSurface from the barrel, so the feature gains a Storybook-only public entry, features/workbar/stories. It cannot share testing: that module is loaded by node --test against tsc output, and the surface and its tool panels use extensionless relative specifiers that only a bundler resolves, so re-exporting it there breaks every node suite behind testing at load time. Stories run through Vite. workbar-boundary.test.ts recognizes the new entry alongside index and testing; nothing that ships imports it.

	before	after
entry chunk	2,245.33 kB (gzip 451.12)	2,166.20 kB (gzip 426.86)
workbar-surface chunk	—	81.31 kB (gzip 25.41)
INEFFECTIVE_DYNAMIC_IMPORT	yes	no

One duplicated authority

assistant-stream and thinking-stream were two implementations of one concept — identical Options/Result shapes, the same redact-append-cap pipeline, the same non-string guard, structurally identical complete paths. They differed only in caps, marker strings, and which end of an over-cap buffer survives, and streaming-display-redaction already parameterizes that as recovery: 'head' | 'tail'.

stream-delta now owns the pipeline. Both modules keep their exported names, cap constants, and option types as thin wrappers. Head-keep's short-circuit on a frozen buffer is now expressed as a property of recovery: 'head', which is what it always was — tail-keep has no fixed point, so it is unaffected.

stream-delta is tested directly, not only through the wrappers: each wrapper exercises one recovery direction, so neither would fail first if the two stopped differing where they must. A mutation that routes head-keep through the tail-keep branch fails the new suite.

tool-output-stream is deliberately left alone: it accumulates a chunk array with dedup-by-seq, a different problem.

Two boundary escape hatches

Four workbar files imported ComposerProps through a seven- or eight-level relative path into packages/ui/dist/composer.d.ts, a build artifact. @maka/ui exports the type and all four already import from it.

Two local simplifications

• chat-message-surface: the first two render-phase activation branches called setActivation with a byte-identical literal — now one condition. The snapshot handed to ChatView also dropped a session guard whose else branch was unreachable: every branch reseeds sessionId to activeSessionId, and a render-phase setState re-runs the body before anything commits.
• use-roving-row-focus: the dependency-array-free layout effect (deliberate, and left that way) wrote tabIndex on every row on every render. An identical write still invalidates style and forces synchronous layout — DevTools charged it 224 ms of reflow. It now compares first. This helps the module pages that use the hook (skills, scheduled tasks, MCP); the chat transcript does not use it.

Verification

• npm run format, npm run check:asf-headers, biome lint over the touched trees
• @maka/ui build + full packages/ui test suite — 256/256, including the new stream-delta tests
• desktop tsconfig.renderer.json and tsconfig.storybook.json typechecks
• apps/desktop/src/main/__tests__/{assistant,thinking}-stream.test.ts — 12/12
• vite build of the renderer, A/B against the same tree with the barrel export restored
• full apps/desktop node suite off build:main — 1640/1640, including workbar-boundary.test.ts

[jackwener]

I reviewed exact head 0309897ef047d7897c5bf57674d32fb156b642c2.

Workbar no longer re-exports WorkbarSurface from the production barrel, so the lazy import can actually split. Assistant and thinking streams share one redact-append-cap pipeline and keep their own caps and head/tail recovery. ComposerProps comes from @maka/ui. Roving-row tabIndex is written only when the value changes.

No blocking issues. One minor note inline: the new shared pipeline has no tests of its own.

---

Posted by an automated review agent operated by @WAWQAQ. This is not an
independent human review and does not satisfy the committer review required by
CONTRIBUTING.md. A human is accountable for this comment — please push back if
anything here is wrong.

简体中文

我审了当前 head。没有阻塞问题。一处可延后：共用流式管道没有自己的测试。

本条评论由 @WAWQAQ 运行的自动化审查程序发出。它不构成 CONTRIBUTING.md
所要求的独立人类审查，也不能替代人类审查。有人类对本条评论负责，如有错误请直接指出。

[jackwener]

Minor: this is the new shared pipeline, but the tests still only go through applyAssistantDelta / applyThinkingDelta. If head-keep and tail-keep later diverge here, those wrappers may not fail first. Not blocking.

简体中文

可延后：共用管道没有直接测试，现在只覆盖两条包装函数。

[Astro-Han]

Addressed the note on stream-delta.ts:119 — added packages/ui/src/__tests__/stream-delta.test.ts, squashed into the pipeline commit.

Nine tests against applyStreamDelta / applyStreamComplete directly, aimed at what the wrappers structurally cannot catch — the points where the two recovery directions must disagree:

• head-keep vs tail-keep from one identical input, both exact strings
• the carried redactionState dropped only where head-keep cut the suffix it describes
• a full head-kept buffer freezing, while a tail-kept window keeps sliding
• the per-delta cap staying identical in both directions
• the non-string guard and state pass-through, both directions
• applyStreamComplete in both directions

Verified the suite actually bites: routing head-keep through the tail-keep branch fails it (8/9), and the tree is back to 9/9 with that mutation reverted.

Note that the approval above is against 0309897ef, which has since been superseded. That head failed CI — re-exporting WorkbarSurface from testing broke every node suite behind that entry at load time, because the surface and its tool panels use extensionless relative specifiers only a bundler resolves. The current head gives Storybook its own features/workbar/stories entry instead. Full apps/desktop node suite is 1640/1640 locally.