refactor(ui): chat message body → chat-message.css + .maka-prose layer (#546 PR4)

[Astro-Han]

Summary

#546 PR4 (first of Phase B). Consolidates the chat message-body CSS into a single per-surface file chat-message.css, then splits the Markdown prose typography off the assistant bubble shell into a reusable .maka-prose element layer that PR5 tool-result bodies will opt into.

Two commits, both zero visual change:

1. Relocate — .maka-message-row, the .maka-bubble-assistant shell + all Markdown prose (p / h* / ul / code / pre / blockquote / table / hr), task-list checkboxes, .maka-code-block*, .hljs*, .maka-bubble-with-actions + .maka-message-copy*, and the .maka-bubble-streaming caret (out of maka-tokens.css @layer components), plus the message-row residue .maka-message-time-inline / .maka-message-meta / [data-slot=message] pre (out of tool-output.css) → styles/chat-message.css. tool-output.css is deleted. Follows the PR1–PR3 unlayered pattern — prose elements are react-markdown HTML with no Tailwind utility competing for them, so the layer move is visually neutral.

2. .maka-prose layer — the shell (.maka-bubble-assistant) keeps only container geometry (max-width 72ch, padding, generic first/last-child margin reset); element typography moves to .maka-prose. bubbleVariants assistant now emits "maka-bubble-assistant maka-prose" so an assistant message renders the prose layer. Same declarations, new scope — PR5 tool-result bodies can reuse .maka-prose without inheriting bubble geometry.

Why

Refs #546. Phase A (PR0–3) dissolved the grab-bag files; Phase B polishes each surface onto its own correctly-named file. PR4 is the chat message body: it had no home of its own (prose lived in maka-tokens.css, meta residue in tool-output.css, body in TSX). .maka-prose is extracted now so PR5's tool-result Markdown reuses the same rules instead of duplicating them.

Scope

Changed: chat-message.css (new), maka-tokens.css (−440 lines), tool-output.css (deleted), styles.css (import swap), primitives/chat.tsx (cva + 2 stale-comment updates), markdown-prose-contract.test.ts (new), 4 existing contracts adapted (visible-copy-hygiene reads chat-message.css; chat-preview-cascade scans the renderer tree instead of the deleted tool-output.css; renderer-css-duplicate-decl tracks multi-line comment state so a prose comment line like max-width: 100% downstream… isn't misread as a declaration; border-width + chat-primitive-cascade follow the .maka-prose rename).

Not included: tool-result reuse of .maka-prose (PR5), chat-view.tsx attachment Tailwind convergence (separate polish), token value changes.

Verification

• Full dist/main suite: 2142 pass, 0 fail (incl. new markdown-prose-contract + all converge contracts).
• npm run typecheck green (all workspaces).
• npm run build:renderer (vite) succeeds — CSS imports + chat-message.css syntax valid.
• turn-narrative screenshot harness: 8/8 variants captured, 0 failed.
• Zero-visual-change reasoning: declarations are byte-identical (relocate copies verbatim; the rename changes selector scope, not declarations), and the @layer components → unlayered move is cascade-neutral for prose because markdown-body.tsx's components map only overrides a/code/pre — every other prose element (p / h* / ul / li / blockquote / table …) is bare react-markdown HTML with no Tailwind utility competing, so unlayered rules win the same way the layered rules did over preflight. Not done: main-baseline byte-for-byte screenshot diff — the only committed turn-narrative baseline predates PR0–3 (2026-07-05 vs HEAD 07-07), so a diff would conflate PR4 with Phase A; the contract + declaration-equivalence evidence above is the stronger guarantee for a pure relocate.

User-facing impact

None — pure CSS relocation + selector rename, zero visual change.

Reviewer notes

• The relocate changes the cascade layer of these rules (@layer components → unlayered), matching PR1–PR3. Intentional and analyzed safe (see Verification).
• .maka-prose is authored as a shared layer but only the assistant bubble consumes it today; PR5 wires tool-result bodies onto it.
• Review focus: the 49 prose selectors renamed in chat-message.css (.maka-bubble-assistant X → .maka-prose X), and that the 4 shell selectors on .maka-bubble-assistant (the container + 3 generic :first/last/nth-last-child resets) are correctly preserved.

Checklist

• Scope matches the PR title and excludes unrelated changes
• Verification lists commands/results
• User-facing impact: none
• Risk/review focus called out (layer move, rename)
• UI changes: screenshot captured (turn-narrative 8/8); byte-diff skipped with justification