fix(crdt): materialize synced deltas into sparse document store (closes #146)

[farhan-syah]

Summary

Fixes the Origin-side root cause of #146: a lite→Origin synced CRDT (Loro) delta was acked and merged into CRDT state (and FTS-indexed via the separate FtsIndex frame) but was never written into the sparse DOCUMENTS btree that DocumentScan / ShapeSnapshot read. As a result a synced collection was catalog-visible but not shape-servable — a plain SELECT id and the ShapeSnapshot both returned 0 rows (doc_count=0).

Change

In execute_crdt_apply (Data Plane), after a delta applies cleanly (both the sync and non-sync arms), the merged Loro row is materialized into the sparse document store:

• encode_crdt_row — reads the merged row, converts LoroValue → JSON → msgpack → canonicalize_document_for_storage, matching the exact bytes the native schemaless put path stores. Runs while the engine borrow is live and returns owned bytes so the borrow drops before touching self.sparse.
• materialize_synced_document — writes to the sparse store keyed by surrogate_to_doc_id(surrogate) (identical to the native path); bitemporal collections use versioned_put, others put. FTS is intentionally not re-indexed here (delivered via the separate FtsIndex frame). Write failures are logged and swallowed so a materialization miss never wedges the sync stream.

Idempotent under replay: non-bitemporal put overwrites by key, bitemporal versioned_put appends one version per applied delta, and duplicate deltas are rejected by sync_admit before reaching apply.

Verification

• 104 crdt/document unit tests green; clippy + fmt clean.
• Cross-repo lite→Origin E2E (in nodedb-lite): a collection created only on lite, synced with no Origin pre-create, is now returned by pg_class, text_match, and a plain SELECT id (the DocumentScan/ShapeSnapshot path).

Follow-ups (out of scope)

The synced-doc apply path does not yet emit CDC WriteEvents, update column stats, or maintain secondary/vector/spatial indexes the way the native write path does.