fix(desktop): keep context gauge stable at the transcript tail

[me2seeks]

Summary

Keep the composer context gauge visually stable when the transcript reaches the live tail.

The tail transition fades the scroll-to-bottom control inside the same filtered dock surface as the composer. Chromium can re-raster that shared surface even though the gauge, model-control group, and composer DOM boxes never move; the thin gauge paths then visibly hop by a device pixel. This patch adds paint containment and a compositing hint to the scroll button’s existing 32px transition container, limiting the opacity repaint to that control.

The existing reader-scroll Storybook scenario now includes a real context gauge and asserts both sides of the contract: the scroll-button paint boundary is present, and the gauge SVG stays within 0.25px while returning to the tail.

Fixes #4973

Verification

• npm --workspace @maka/desktop run typecheck:stories
• node --test packages/ui/dist/__tests__/composer-context-usage.test.js (2/2)
• npm --workspace @maka/desktop run build-storybook
• node scripts/storybook-visual-smoke.mjs apps/desktop/storybook-static (312 stories, 339 theme renders)
• npm run check:renderer-architecture -- --base origin/main (101/101 fixtures; check passed)
• npm run astryx:surface-inventory
• npm run lint
• npm run format:check
• git diff --check

AI use

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool and scope: Codex — diagnosis, implementation, browser verification, and test coverage.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck, and affected suites pass locally

Does this PR entail a change in behavior?

• Yes — the context gauge no longer visually jumps during the tail transition
• No

[jackwener]

Approving at exact head 6572edb40. test is terminal green.

One thing is worth stating plainly, because the title reads the other way: this change does not touch the gauge's data flow at all. The diff is twelve lines of CSS giving the scroll-button's child its own paint and compositing boundary, plus a Storybook assertion. The production chain still re-reads the latest settled context snapshot and calls onChange, and the focused tracker test exercises an at-tail update from 40,000 to 52,000 tokens.

So "stable" here means visually stable, not frozen. That distinction was the thing worth checking: a gauge that stops flickering because it stopped updating would look like a fix and be a worse bug, harder to notice than the flicker it replaced. It is not that — the update path is untouched, which is the strongest form this fix could take.

The comment in the CSS is the most valuable part of the change. It names the mechanism — during the opacity transition both controls share the dock's filtered compositing surface, so Chromium may re-raster the thin gauge paths and the icon visibly hops by a device pixel although its DOM box does not move. A future reader tidying up will-change and contain will know why the rule exists, which is what usually gets lost.

Narrower check coverage here than on the sibling pull requests: only test ran on this commit, where others also ran label. Not a defect — context for how much the green covers.

This is not the independent human review CONTRIBUTING.md requires, and it is not a merge decision.

简体中文

在 exact head 6572edb40 上批准。test 已终态转绿。

有一点值得明说,因为标题读起来是另一个意思:这次改动完全没有触碰计量的数据流。 diff 是十二行 CSS,给滚动按钮的子元素一个自己的绘制与合成边界,外加一条 Storybook 断言。生产链路仍然重读最新的已结算上下文快照并调用 onChange,而定向 tracker 测试覆盖了尾部 40,000 到 52,000 tokens 的更新。

所以这里的「stable」指的是视觉上稳定,不是被冻住。 这个区分正是值得去查的东西:一个因为不再更新而不再闪烁的计量,看起来像修复,实际是更糟的 bug —— 比它所取代的闪烁更难被发现。它不是那样 —— 更新路径未被触碰,这是这个修复所能采取的最强形式。

这次改动里最有价值的是 CSS 里那段注释。 它点明了机制 —— 在透明度过渡期间两个控件共享 dock 的滤镜合成面,于是 Chromium 可能重新栅格化细线计量路径,图标在 DOM 盒子没有移动的情况下视觉上跳一个设备像素。日后有人整理 will-change 与 contain 时,会知道这条规则为什么存在 —— 而这恰恰是通常会丢失的东西。

这里的检查覆盖比同批其他 PR 更窄: 该 commit 上只跑了 test,而其他几个还跑了 label。不是缺陷,而是「这片绿覆盖了多少」的背景。

这不是 CONTRIBUTING.md 所要求的独立人类审查,也不是合并决定。

---

Automated review notice: This comment was posted by an automated review agent operated by jackwener. It is not an independent human review and does not replace one.