fix(desktop): stop the transcript boundary from clipping the first glyph

[Adarsh-Me]

Summary

Fixes #4898. content-visibility: auto on [data-maka-transcript-boundary] (added by d2d9f90 for #4259) applies paint containment, which clips descendants at the marked element's own box. The ghost assistant bubble carries no horizontal padding (padding: var(--space-0-5) 0 0), so the containment clip edge sat on the line-box origin itself: glyph ink reaching a pixel or two past the origin — italic overhang, CJK ideographs — lost its leftmost column, and the first character of each assistant line rendered partially.

The fix keeps the containment (and with it the #4259 rendering bound) and moves the clip edge instead of the text: the assistant boundary gains padding-inline: var(--space-1) with a matching margin-inline: calc(-1 * var(--space-1)). The clip boundary now sits four pixels past the text on each side; the text keeps its exact position; the ghost bubble paints nothing of its own, so the wider box is invisible. Card-styled boundary sites (reasoning, tool activity) already pad their content, so they are untouched.

Verification

Measured live on the fixture the issue names (product-shell-official-appshell--partial-history-notice, served by Storybook from this branch):

• Before (inline overriding the fix off): containment box left edge at x=316, first glyph origin at x=316 — zero clearance; ink at the boundary itself is clipped, which on the issue's macOS/zh-CN rasterization visibly cut the first stroke.
• After (rule applied): containment box left edge at x=312, first glyph origin at x=316 — 4px of guaranteed clearance, text at the identical visual position.
• Screenshots of the fixture before/after confirm the rendered output is otherwise unchanged.

AI use

Select exactly one:

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

Tool(s) and scope:

Checklist

• Tests cover the change and fail without it — the failure mode is compositor-side (paint-containment clipping); jsdom cannot see it, so the differential was verified against the running Storybook fixture with before/after measurements instead of an automated test. Reproduction steps from the issue apply directly: without this rule the first glyph origin sits at the clip edge with zero clearance.
• Lint, format, typecheck and the affected suites pass locally — biome check clean on the changed file; CSS-only change, no TS surface touched.

Does this PR entail a change in behavior?

• No — clipping behavior only; layout position, containment perf bound, and scroll anchoring are unchanged.

[Astro-Han]

Thanks. Closing in favor of #4906.

Head b7c6c1e. The diagnosis matches #4898: content-visibility: auto from #4259 paint-clips at the boundary box. The padding-inline + negative-margin compensation is the wrong tool for that. It moves layout to dodge a paint clip.

CI already shows the cost. Storybook smoke failed on product-shell-official-appshell--wide-assistant-prose: expected 8 to be less than or equal to 1. That story is the reading-column flush contract. Negative margin does not buy width against the parent clip, so the 4px padding just insets the prose.

#4906 keeps containment and sets overflow-clip-margin on the existing boundary rule, which is the clip edge. CI is green there.

Reviewed with Grok.