perf: isolate chat streaming renders and reduce sidebar render cost (#1672)

Reworks the chat and session-sidebar render paths to cut render cascades, memory
  churn, and UI jank on large sessions and big session trees. Behavior is preserved;
  the changes are about *when* and *how much* the UI re-renders.

  ## Chat streaming
  - Freeze the streaming message's parts in the bulk turn projection during streaming,
    and re-inject live parts only in an isolated tail leaf, so a ~60/sec delta stream
    no longer re-runs the whole-session projection or re-renders unrelated rows.
    session with referential reuse of unchanged turns.
  - Memoize message rows with field-aware comparators instead of reference equality.
  - Replace the manual child-session polling in the task tool with the live SSE
    stream + a one-shot load, removing a fetch/settle state machine.

  ## History loading & scroll
  - Load an initial page fast, then prepend one older page in the background so the
    scroll container has headroom and "load older on scroll-up" fires before the user
    hits the absolute top.
  - Compensate scroll synchronously (in a layout effect, before paint) for prepends —
    including background prepends that don't originate from a user scroll — so the
    viewport stays stable instead of judder-correcting on the next frame.

  ## Markdown rendering
  - Render markdown synchronously *styled* on first paint (paragraphs, lists, code
    cards, tables, inline code) instead of raw escaped text; the async pass then only
    upgrades syntax-highlight colors. Eliminates the flash of full-width raw text.
  - Load KaTeX CSS eagerly with the main bundle instead of inside the lazy markdown
    chunk, avoiding a late stylesheet injection on first render.

  ## Sidebar
  - Hoist per-row recursive tree walks out of row comparators into per-group
    precomputed sets/keys; batch live-session lookups into a single map; add a
    group-level memo boundary.
  - Isolate rename drafts so per-keystroke typing doesn't repaint the row tree.

  ## Sync layer
  - Add a staleness guard so a slow message fetch can't repopulate a session the user
    navigated away from.
  - Throw on fetch failure for authoritative loaders so a transient blip can't read as
    an empty server response.

  ## Cleanup
  - Remove dead code (unused hooks, params, duplicated inline types) surfaced while
    reworking the above.

  ## Known issue
  - A rare, purely cosmetic first-paint width flash can still appear on large sessions;
    it has no behavioral or data impact and is tracked for a follow-up runtime trace.

packages/ui/src/styles/typography.css

@@ -158,3 +158,18 @@
     font-weight: var(--markdown-h4-font-weight);
   }
 }
+
+/* Chat message column width — intentionally UNLAYERED so it always wins the
+   cascade. The same rule inside `@layer components` above is correct in steady
+   state, but a layered rule can be momentarily out-ranked during a heavy style
+   recalc (e.g. a large session's virtualized list re-laying-out when the
+   background history prepend lands, or a late-loaded CSS chunk re-establishing
+   layer order). That dropped the clamp for a frame and flashed message text to
+   full width. An unlayered duplicate of just the width keeps it robust. */
+.chat-message-column {
+  width: min(100%, 48rem);
+}
+
+.wide-chat-layout .chat-message-column {
+  width: min(100%, 64rem);
+}