Render the assistant's replies as Markdown

The assistant answers in Markdown — tables, lists, bold, links — but the
chat bubble showed the raw source (whitespace-pre-wrap), so a table came
through as a wall of pipes.

- New MarkdownMessage: react-markdown + remark-gfm (for GFM tables). It does
  NOT enable raw HTML (no rehype-raw), so a model reply can't inject markup;
  every element the assistant actually uses is Tailwind-styled for a chat
  bubble, and wide content (tables, code) scrolls in its own box so the
  bubble can't blow out.
- Only ASSISTANT bubbles render as Markdown; the user's own text stays
  literal (their `*` shouldn't become a bullet) and the assistant bubble goes
  full-width so a table has room.
- Lazy-loaded: the renderer + its ecosystem (~150 KB) is its own async chunk,
  loaded only when an assistant message renders — not for everyone who opens
  the editor. To keep it out of the eager first paint, manualChunks now
  vendors only the app-wide core (react/react-dom/scheduler kept together —
  splitting react-dom breaks React 19 at load) and lets everything else ride
  with its importer.

Tested via renderToStaticMarkup (node, no DOM): a GFM table renders with
styled cells + horizontal scroll; **bold**/lists render; raw <script>/<b> are
escaped not emitted; links get rel="noopener noreferrer". App re-verified to
mount cleanly with the re-chunk (no React-19 init-order break). tsc + vitest
(99) + build green, no >500 KB warning.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01H3zbym8Doka2d7D48maSgZ

web/package.json

@@ -22,6 +22,8 @@
     "clsx": "^2.1.1",
     "react": "^19.0.0",
     "react-dom": "^19.0.0",
+    "react-markdown": "^9.1.0",
+    "remark-gfm": "^4.0.1",
     "tailwind-merge": "^2.6.0",
     "zod": "^3.24.1",
     "zustand": "^5.0.14"