[cyberspace-cs]

Summary

Fixes #556.

DatabaseManager.batch_compare_queries previously validated every distinct db_id up front and, when any one was missing, returned an error result for all triples:

if not self.database_exists(db_id):
    return [self._create_error_result(error_msg) for _ in query_triples]

Because SQLExecutionClassifier submits the whole dataframe in one call, a single row pointing at a missing database made every other comparison fail with "Predicted SQL failed; Ground truth SQL failed", corrupting the difficulty labels for the entire dataset — even rows whose database exists and whose SQL matches.

Fix

• Validate each triple individually.
• Execute only the triples whose database exists (in a single batched batch_execute_queries call, as before).
• Emit a per-row error result for missing databases, and place results back at their original indices so output order always matches input order.
• Empty batch still returns []; existing equal / unequal comparison behavior is unchanged.

Behavior

For the reproducer in #556:

Triple	Before	After
(concert, same sql, same sql)	❌ False, Database 'missing_db' not found	✅ True, []
(missing_db, ...)	❌ error	❌ False, Database 'missing_db' not found

Tests

Added test/cpu_only/test_batch_compare_queries_missing_db.py (cpu-only, SQLite, no model/network):

• a missing db only fails its own triple, in both orderings
• each missing db gets its own error message
• valid but unequal queries are still compared
• empty batch returns []

pytest -m cpu test/cpu_only/test_batch_compare_queries_missing_db.py → 5 passed.

[cyberspace-cs]

Closing this as a duplicate — I missed #557 when I opened it, and it fixes the same root cause (per-row isolation, original order preserved) with the same approach. Two PRs for one issue only splits review attention, so #557 is the one that should land.

If the extra cases in my branch are useful (missing database in the first position, and a batch where every database is missing), I am happy to port them into #557 as a follow-up commit rather than keep this open. Sorry for the noise.