The Tauri relay (NotebookSyncClient) still holds a full AutoCommit Automerge replica. With daemon-owned loading (#598) complete, the relay's doc is only needed for:

1. get_doc_bytes() — frontend fetches full doc bytes to seed its WASM handle
2. SyncUpdate metadata diffing — detects metadata changes to emit notebook:metadata_updated events

3. get_cells() in save_notebook_as — removed in PR #602
4. Cell population via initialize_notebook_sync — removed in PR #602

Remaining work

Reroute get_doc_bytes()

The frontend calls invoke("get_automerge_doc_bytes") at bootstrap to seed the WASM handle. Currently this reads from the relay's AutoCommit. Options:

• Daemon request: Add a NotebookRequest::GetDocBytes that returns the daemon's canonical doc bytes. The relay forwards the response without parsing.
• WASM export: Have the frontend WASM export its own bytes. But on first bootstrap, the WASM doc is empty — it needs bytes from somewhere.

Daemon request is the cleaner path.

Move metadata change detection to frontend

The relay's run_sync_task uses SyncUpdate.notebook_metadata to detect metadata changes and emit notebook:metadata_updated. The frontend WASM doc already has full metadata — when receive_sync_message applies changes, the frontend can diff and emit events locally.

Strip the relay

• Remove doc: AutoCommit from NotebookSyncClient
• Reduce SyncCommand enum — remove AddCell, DeleteCell, UpdateSource, AppendSource, ClearOutputs, AppendOutput, SetExecutionCount, GetCells, SetMetadata, GetMetadata, GetDocBytes
• Keep: SendRequest, ReceiveFrontendSyncMessage, and lifecycle commands
• Keep peer_state + frontend_peer_state — Automerge sync protocol still needs peer state tracking

runtimed-py is unaffected

The Python SDK uses connect_split (not connect_split_with_raw_sync) and legitimately operates as a full Automerge peer. The simplification is Tauri-path-specific.

Impact

Eliminates 8 correctness findings from the protocol audit (including 2 High severity):

• sync_to_daemon() broadcast dropping
• changed_rx.recv() Lagged/Closed handling
• Triple-merge divergence risk
• try_send silent drops on changes_tx
• biased select starvation
• Virtual sync handshake 10-iteration limit
• raw_sync_tx send failures
• receive_and_relay_sync_message wrong peer state

Size

L (but closer than originally scoped — cell population and get_cells() already removed)