Test shinychat's `history=` feature instead of ad hoc bookmark state

Background

shinychat 0.7.0 added a first-class conversation-history feature: Chat(history=...), either True/False or a HistoryOptions instance with restore_mode ("browser" / "url" / "none" / "bookmark"), store ("auto" / "memory" / "file" / a custom ConversationStore), scope, title and max_store_mb.

We have no coverage of it. What we have instead is tests/playwright/shiny/bookmark/chat/client_state/, which predates the feature and hand-rolls the same job: a RepeaterClient implementing get_state/set_state, wired up through chat.enable_bookmarking(client, bookmark_store=...).

Two problems with that as our only coverage:

1. It exercises plumbing an app shouldn't need to write any more. restore_mode="bookmark" participates in Shiny server bookmarking directly, and restore_mode="url" persists conversations with no Shiny bookmarking at all.
2. It doesn't touch the parts most likely to break — the conversation store, scoping, restore modes, switching between conversations.

Why now

The app passed messages= to the Chat constructor, which shinychat 0.7.0 turned into a hard error (Chat(messages=...) requires history=False), so these tests went red on main for days. #2486 migrates the startup message to greeting= and points the assertion at expect_greeting, which gets CI green and incidentally covers the greeting bookmark round-trip — but it leaves the ad hoc state plumbing as the thing under test. There's a TODO in the test file pointing here.

What to cover

• history=True vs False, and the history= + messages= interaction that now raises.
• Each restore_mode: "browser" (localStorage, survives reload without touching the URL), "url" (?shinychat_conversation_id=), "none", "bookmark" (requires bookmark_store="server").
• store="memory" — the docstring calls it out as the testing option.
• scope for multi-user separation.
• Switching conversations in-session, and @chat.history.on_save / on_restore callbacks.

Note: history= only takes effect when client= is also provided, so most of this needs a chatlas-shaped client. That may mean a fake client implementing the protocol rather than a real LLM.

Start in tests/playwright/shiny/bookmark/chat/client_state/ and follow the TODO, then inspect shinychat's Chat(history=...) and the available chatlas-shaped client protocol. Run the relevant Playwright tests while covering history settings, restore modes, stores, scoping, conversation switching, and callbacks; done means the first-class history feature replaces the ad hoc bookmark-state coverage.