perf: reduce memory usage with lazy schema loading and tab eviction

[datlechin]

Summary

• Replace bulk fetchAllColumns() with lazy per-table column fetch and LRU cache (max 50 tables) in SQLSchemaProvider — avoids loading all column metadata upfront for large databases
• Evict inactive tab row data more aggressively: keep at most 2 inactive tabs loaded, include both table and query tabs, exclude tabs with unsaved changes
• Defer Phase 2 background work (exact COUNT + enum values) by 200ms so Phase 1's QueryResult can be deallocated first, reducing peak memory overlap
• Share the coordinator's SQLSchemaProvider with SidebarViewModel so sidebar reads cached tables instead of issuing a redundant fetchTables() query
• Release QueryResult earlier in executeQueryInternal by wrapping extraction in a do {} block before the await MainActor.run suspension point

Test plan

• 8 tests for lazy schema column loading + LRU eviction (SQLSchemaProviderTests)
• 7 tests for tab eviction filtering and budget logic (TabEvictionTests)
• 4 tests for sidebar schema provider cache integration (LiveTableFetcherTests)
• All 19 tests pass on arm64 and x86_64
• Build succeeds with 0 lint violations
• Manual: connect to a database with 50+ tables, verify autocomplete still works
• Manual: open 5+ tabs, verify switching re-fetches evicted data correctly
• Manual: verify Phase 2 data (exact row count, enum dropdowns) still appears after brief delay

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.