Fix race: client receives broadcast updates before sync handshake completes

[xrl]

Problem

YRoom.serve() adds the client to self.clients immediately upon connection — before the SYNC_STEP1/SYNC_STEP2 handshake finishes. If the server-side document is mutated concurrently, _broadcast_updates sends a SYNC_UPDATE to the un-synced client. The client-side yjs then throws "Unexpected case" in findIndexSS because the update references a struct at a clock the client doesn't have yet.

How to trigger

This happens reliably in JupyterLab with jupyter-collaboration when an AI agent (via MCP tools) rapidly adds cells to a notebook while a new browser tab opens the same notebook. The new tab's WebSocket connects and starts the sync handshake, but before SYNC_STEP2 is received, a broadcast update arrives referencing a struct one clock tick past what the client has.

Debug data from instrumented yjs findIndexSS

[yjs findIndexSS] FAILED — clock not found in structs
{ searchClock: 1223, structsLength: 75, allRanges: "0..0, 1..8, ..., 1200..1222" }

The client has structs through clock 1222. The server sent an update referencing clock 1223. Always exactly one tick past the end — one broadcast update leaked through before the sync completed.

The race

1. Client connects → self.clients.add(channel)     ← TOO EARLY
2. Server sends SYNC_STEP1 (state vector) to client
3. MCP add_cell mutates server doc → _broadcast_updates fires
4. _broadcast_updates sends SYNC_UPDATE to ALL self.clients
   (including the new, un-synced client)
5. Client receives SYNC_UPDATE referencing clock 1223
6. Client hasn't processed SYNC_STEP2 yet — only has through 1222
7. findIndexSS throws "Unexpected case"

Fix

Defer self.clients.add(channel) until after the first sync message exchange completes. The client only joins the broadcast list once it has a complete initial state.

Also changes self.clients.remove() to self.clients.discard() in the finally block, since a client that disconnects before syncing was never added to the set.

Test

Added test_broadcast_not_sent_before_sync_completes which:

1. Creates a YRoom with existing data
2. Connects a new client
3. Mutates the server-side doc 20 times during the handshake
4. Asserts the client receives all data without errors

Without the fix, the client would crash on "Unexpected case" during step 3.

[davidbrochart]

I think that we are synchronized when we receive a SYNC_STEP2 message:

Suggested change

	if not synced:
	if not synced and message[1] == YSyncMessageType.SYNC_STEP2:

[davidbrochart]

Sorry, that's not correct. The client is synchronized when the server sends a SYNC_STEP2 after receiving the client's SYNC_STEP1, so we don't need the synced variable:

Suggested change

if not synced: