[httpsphl]

Third slice of the performance round. It is independent of #75, and each helps on its own.

What changes

• Memo on BotRun (compares the items one by one; withItem keeps unchanged items stable), InboundRow, NoticeRow, ChatComposer and ChatView. Each chat.delta now re-renders only the run being written, not every run, tool line and card in the chat.
• chatRows runs only when the items change. useFiles and onSent keep their identity.
• BotView's showFile / showBrowser / showScreen are stable (ui/useStable.ts), so the chat skips BotView's re-renders: bot changes, and screen.draft every 250 ms.
• The draft is split after its last finished block (never inside a code fence, never before an indented line). The finished part is parsed once, and only the tail again as text arrives. Before, the whole draft was re-parsed on every piece.
• day() and decimal() reuse their Intl formatters per language.

Not done here, for now: batching chat.delta by requestAnimationFrame in the app. #75 already sends at most ~25 per second per bot.

Tests

• stream.test.tsx counts the renders of a finished run's ToolLines while text streams: 0 extra, against 4 without the memo (I checked that it fails).
• rows.test.ts: splitDraft with paragraphs, a code fence and an indented line.
• pnpm check passes. The FakeBotloft preview shows the chat as before.

Manual check (real Claude)

• Ask a bot for a long answer with paragraphs, a list and a code block. The text appears normally while being written, the code block does not break mid-way, and the final reply replaces the draft without a jump.