aditya3799
### Background In a previous optimization (PR #19), a redundant call to `selected_or_default_graph_id_via_definer()` was successfully removed from `current_catalog_state()`. This helped reduce the ~3.5ms unamortized fixed cost of graph traversals. ### The Problem However, looking closely at `ensure_current_graph()`, the very next line calls `pending_sync_rows(applied_sync_id)`. Tracing this call shows that it eventually hits `SyncReplayContext::load()`, which triggers a full `read_catalog()`. This means `ensure_current_graph()` is *still* doing a full redundant catalog read (including two massive SPI queries for `_registered_tables` and `_registered_edges`), completely undermining the optimization from PR #19. ### Proposed Solution Since `ensure_current_graph()` already loads `catalog_state` right before calling `pending_sync_rows()`, we can extract the `applicable_table_oids` from the loaded state and pass them directly to `pending_sync_rows`. This bypasses `SyncReplayContext::load()` entirely and eliminates the remaining redundant SPI queries from the hot path. I will submit a PR with this fix shortly!