perf(chat): virtualize the agent-chat message list

[Zeus-Deus]

Closes #77

Summary

Replaces the plain slots.map() transcript render in MessageList.tsx with a virtualized list, so only the on-screen window of rows is mounted regardless of conversation length. A 797-message session now mounts at most ~20 row nodes (was: one DOM subtree per message, up to the 5,000-message reducer cap).

Library choice (issue item 1)

react-virtuoso v4.18.7 (MIT). Evaluated against @tanstack/react-virtual (MIT, headless — would require hand-rolled dynamic measurement + bottom-pinning) and react-window (MIT, weak variable-height support). Virtuoso is the only one with built-in ResizeObserver-driven row measurement and bottom-anchoring primitives — exactly the two hard parts called out in the issue. The commercially licensed @virtuoso.dev/message-list package is not used.

What's preserved (issue item 2)

• Stable keys + memo rows — per-slot keys are still slot.item.id / run:<first-id>, rows render through MessageRowMemo; a streaming token re-renders exactly one row.
• Tail-follow auto-scroll — MessageList owns the scroller now (the virtualizer must control it). Pinned-ness is tracked from real scroll events (≤ 80 px, same threshold as before); every transcript change snaps to the tail only while pinned. Content growth never fires a scroll event, so streaming can never unpin — and a user wheel-up unpins immediately, so auto-scroll never fights the reader.
• Variable row heights — tool-run collapses render as a single virtual row that expands in place; Virtuoso re-measures via ResizeObserver. The thinking pulse renders as the last virtual row (not a footer) so the tail snap keeps it visible while streaming.
• Tool-run collapse / merged-slot logic — unchanged; ChatTranscript shrinks to a thin shell that derives showThinking.

Also in this PR

• StrictMode hydrate fix (AgentChatPane.tsx): hydrate-on-mount never applied in dev builds — the first effect invocation was cancelled by StrictMode's mount/unmount cycle and its attempt marker blocked the re-run. Production (single mount) was unaffected.
• Dev-mock chat harness (src/dev/): seeded agent-chat-demo workspace with enable_agent_chat on, a ~790-row transcript hydrated through the real reducer via agent_chat_list_messages, simulated streaming replies on agent_chat_send_turn, and window.__codemuxChatMock.streamReply() for on-demand streams. Verified tree-shaken from the production bundle.
• Docs: virtualization contract + dev harness documented in docs/features/agent-chat.md.

Acceptance criteria — verified E2E in the real UI (browser mock, real store/reducer/event pipeline)

• 1,000+ message session keeps only on-screen rows in the DOM — 797 store messages → max 20 mounted rows across a full-history scroll sweep (11 at rest); session opens at the tail (Turn 220 visible).
• Streaming auto-scroll + manual scroll-up behave — pinned 100-token stream: distance-from-bottom samples all ≤ 80 px, tail visible; mid-stream wheel-up: scrollTop frozen exactly while tokens kept appending; scrolled-up reading position unmoved by a background stream.
• Tokens re-render exactly one row — MutationObserver during a 30-token stream: only the tail item mutated (plus a render-count unit test in MessageList.virtualization.test.tsx).
• Tool collapses, approvals, plan blocks render and resize correctly — collapse expands 4→8 rows in place and back with zero scroll drift; dispatch tests (plan / user-input / generic approval) all green under VirtuosoMockContext.

Verify

• npm run verify fully green: cargo check, 1684 Rust lib tests + all integration suites, tsc --noEmit, 124 test files / 1826 frontend tests (2 new virtualization tests; existing MessageList tests updated to render inside VirtuosoMockContext).
• npm run build (production) passes; dist/ contains no mock code.
• Note: two pre-existing Rust tests are environment-sensitive on dev machines (project_codemux_entry_is_filtered_out reads the real ~/.claude.json; resolve_binary_finds_native_binary_from_project_root depends on a system-installed agent-browser outside npm run PATH). Both pass under a clean-HOME npm run verify; this PR touches no Rust beyond zero lines.

Manual repro for reviewers

1. npm run dev → codemux browser open http://localhost:1420
2. Open the agent-chat-demo workspace — long transcript opens at the tail.
3. DevTools: document.querySelector('[data-testid="virtuoso-item-list"]').children.length stays ~10–20 while scrolling.
4. window.__codemuxChatMock.streamReply('thread-mock-chat', { tokens: 120 }) — pinned: view follows; scrolled up: view stays put.