- Add 150ms debounce on async completion path to reduce redundant Tauri IPC calls during rapid typing
- Parallelize schema table fetching with max concurrency of 5
- Separate schema list cache from table completion cache for longer-lived reuse
- Return keyword completions synchronously when no backend data is needed
- Lower snippet base boost for label-only matches so keywords rank above snippets when user types past snippet prefix