[ZhuchkaTriplesix]

Closes #1572.

Before (measured)

• ERD neighbourhood (ErdCatalog.loadNeighbourhood): for every schema in the neighbourhood it ran the column catalog query over all tables of that schema and kept the few it needed in Dart. On PostgreSQL that query has per-column correlated subqueries (unique index, enum labels, comments, defaults), so its cost follows the size of the schema. With catalogRowLimit = 200000, a large schema was cut before filtering, and the table the user opened could come back without columns.
• MCP describe_table / sample_rows: to find one table, three catalog queries over the whole schema (columns, keys, view columns) on every call. sample_rows used them only to check the name.

SQLite (no PostgreSQL here; the app's own SQL run with Python's sqlite3), tables with 17 columns:

tables	whole-schema columns (before)	filtered to the 3 tables of a neighbourhood (after)
500	35 ms, 8,499 rows	0.3 ms, 51 rows
3,000	225 ms, 50,999 rows	0.7 ms, 51 rows

On PostgreSQL the gap should be wider (the subqueries run per column), and the rows also travel to the app and get decoded.

Change

• columnsSql / viewColumnsSql(…, tables:) add AND <name> IN ('a', 'b') (c.relname / c.table_name / m.name), literal-escaped like the rest of the catalog SQL. An empty list matches nothing; null is the old whole-schema query.
• loadNeighbourhood: keys first (unchanged; the walk needs them), then for each PostgreSQL schema only its tables of the neighbourhood. schema.table asks that schema for table, which also gives the fallback for a key-less table its bare name.
• MCP: tableNamesSql (names of tables and views only) resolves the name with the same rule as before (exact, else the single case-insensitive match). describe_table then loads that table's columns and view columns; sample_rows needs only the name. list_tables and the schema:// resource still read the whole catalog, as they should.

Tests

• erd_test.dart:
  • the neighbourhood query carries the IN filter;
  • a schema whose whole-schema query is truncated before the target still gives the target's columns (and is no longer marked truncated);
  • the filter is emitted for each dialect, quotes are escaped, an empty list matches nothing;
  • existing fakes recognise the query with or without the filter.
• mcp_query_service_test.dart: describeTable sends only filtered column queries, and sampleRows('USERS') resolves users without any column query. The fakes now answer the names query.

Not run with flutter test locally; CI decides.