[raams95]

Summary

DatabaseManager.batch_compare_queries returns an error result for every triple in the batch as soon as any one db_id is not registered. One row that points at a missing database makes all the other comparisons fail too, even when their database exists and the SQL matches.

SQLExecutionClassifier passes the whole dataframe in a single call, so in practice one bad db_id marks every row as "Predicted SQL failed; Ground truth SQL failed" with cnt_true = 0.

manager.batch_compare_queries([
    ("concert", "SELECT name FROM singer", "SELECT name FROM singer"),
    ("missing_db", "SELECT 1", "SELECT 1"),
])
# both results: equal=False, ["Database 'missing_db' not found"]

Fix

The missing-database check now happens per triple. A triple whose db_id is not registered gets the same "Database '<id>' not found" error result as before. The other triples are executed and compared as usual, and results stay in input order. If every database exists, the behaviour doesn't change.

Tests

I added test/cpu_only/test_text2sql_compare_missing_db.py. It builds a small SQLite database under tmp_path and compares three triples: a matching pair, a pair on a missing database, and a non-matching pair. It checks that only the missing-database triple gets the error and the other two keep their real results in order.

• Before the fix: 1 failed (the matching pair came back equal=False)
• After the fix: 1 passed

pytest -m "cpu or api" test/cpu_only/test_text2sql_compare_missing_db.py -q
1 passed

I also checked SQLExecutionClassifier.execute_model_batch with the same data. The row on the existing database now gets cnt_true = 1 instead of being reported as failed.

Fixes #556