What does this PR do?

Adds automatic bidirectional text direction to the desktop chat, so Arabic/Hebrew/Persian/Urdu — and the common mixed RTL/English technical case — read and align correctly instead of running left-to-right with mangled punctuation order.

Each chat block resolves its own base direction from its first strong character (UAX#9), purely in CSS, scoped to the chat surfaces only:

• unicode-bidi: plaintext + text-align: start on assistant prose blocks (p, h1–h6, li, blockquote), the user bubble's text lines, and both composers (main + edit share the composer-rich-input slot). RTL blocks read and right-align RTL; English stays LTR; mixed conversations resolve per block. The text-align: start is load-bearing — the user bubble hardcodes text-left (USER_BUBBLE_BASE_CLASS), so without it RTL prose would read right-to-left but cling to the left edge.
• Inline code and KaTeX are pinned direction: ltr; unicode-bidi: isolate. isolate makes the bidi first-strong heuristic treat each as a single neutral, so a block that starts with a command (`./run.sh -v` followed by Arabic) still resolves RTL, and the command's own neutrals (dots/slashes/dashes) keep their order.
• Fenced code surfaces (code-card, user fences) are pinned LTR so they never mirror or right-align inside an RTL list item or blockquote.

direction is never set, so app chrome, layout, and list indent stay LTR (the issue explicitly asks not to flip the whole UI). List markers stay on the left by design. English-only content is byte-for-byte unchanged.

Supersedes

This supersedes #44065 and #44169 and unifies them into one minimal change (credited via Co-authored-by):

• feat(desktop): per-message automatic RTL/bidi text direction #44065 (feat/desktop-rtl-bidi) took the JS route — a text-direction.ts first-strong resolver plus dir-attribute plumbing across markdown-text.tsx, user-message-text.tsx, thread.tsx, and the composer (~280 lines) — specifically to handle the "code-first" paragraph (a block starting with a command shouldn't flip LTR). I verified in Chromium (Electron's engine) that unicode-bidi: isolate already removes inline code from the paragraph's first-strong vote, so that case is handled in CSS and the JS is unnecessary.
• feat(desktop): auto-detect RTL paragraph direction in chat #44169 (fix/44150-chat-rtl-bidi) was the right pure-CSS approach but didn't text-align: start (so RTL user bubbles read RTL but stayed left-pinned under the bubble's text-left) and didn't pin fenced code blocks against mirroring inside RTL list items.

This PR keeps #44169's CSS-only footprint and fixes both gaps. Please close #44065 and #44169 in favor of this.

Related (out of scope, no overlap)

• Web dashboard: Arabic text in chat pane renders LTR with disconnected letters (no BiDi / no shaping) #29047 — same symptom in the web dashboard chat pane
• feat(tui): RTL/Bidi support for Persian, Arabic, Hebrew, and Urdu text #41462 — RTL/bidi for the TUI
• feat(web/i18n): add Arabic (العربية) with full RTL layout support #29795 / web dashboard Arabic locale — locale work, not direction
• Dashboard /chat is an xterm.js terminal; xterm has no upstream bidi support, so it can't be fixed at this layer.

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

How to Test

1. Run the desktop app, send an Arabic message such as مرحبا كيف حالك — the bubble right-aligns.
2. Ask for an Arabic answer with a heading, a list, and a fenced code block — prose/headings/bullets right-align; the code block stays LTR and un-mirrored even inside an RTL list item.
3. Send a message that starts with a command, e.g. `./scripts/run.sh -v` ماذا يفعل هذا الأمر؟ — it still right-aligns (code doesn't vote on direction) and the chip keeps its internal order.
4. Type Arabic in the composer — it right-aligns as you type; switch to English and it flips back. Same for the edit composer.
5. Send English messages — rendering is identical to main.

Verification

• npm run typecheck — clean
• eslint on touched files — clean
• Empirically validated the bidi behavior (alignment + isolate skipping the code-first vote) in Chromium before implementing.
• Desktop vitest failures on this branch are identical to clean main (pre-existing, unrelated).

Fixes #44150