Guard LLM-generated sql_query rules

[aarushisingh04]

Changes

adds SQL safety validation to both LLM-assisted rule generation paths so that any sql_query rule containing unsafe SQL (DML/DDL) is dropped before being returned to the caller.

• added _filter_unsafe_sql_rules() in llm/llm_core.py : parses the generated JSON rules array and drops any rule whose check.function is sql_query and whose query argument fails is_sql_query_safe(), logging a warning with the sanitized query string (CWE-117)
• hooked it into DspyRuleGeneration.forward() and DspyRuleUsingDataStats.forward() via an else clause on the existing JSON-validation block; DspyRuleGenerationWithSchemaInference inherits the fix automatically since it delegates to DspyRuleGeneration

note: max_tokens was already present in LLMModelConfig and forwarded to dspy.LM(), it caps completion length, not prompt input, which is sufficient for the issue's intent of bounding unbounded LLM output.

• fixed training example for ai-assisted rules generation

Linked issues

resolves #1123

Tests

• manually tested
• added unit tests
• added integration tests
• added end-to-end tests
• added performance tests

19 unit tests in tests/unit/test_llm_core.py covering: safe SELECT queries pass through, unsafe DROP/DELETE/INSERT/TRUNCATE rules are dropped, non-sql_query rules are unaffected, mixed safe+unsafe batches, null/missing query argument, non-array JSON input, malformed rule entries, invalid JSON passthrough, and warning log emission.

Documentation and Demos

• added/updated demos
• added/updated docs
• added/updated agent skills

[mwojtyczka]

Thank you for the PR. Generally looking good. Left some comments to address.

[aarushisingh04]

@mwojtyczka

appreciate the thorough review! addressed all three points in the latest push:

• added isinstance(query, str) guard, non-string query is now dropped with a warning instead of crashing
• refactored _filter_unsafe_sql_rules to accept a pre-parsed list, eliminating the double-parse
• added missing tests: non-string query, non-dict arguments, DspyRuleGenerationWithSchemaInference path, and empty/None quality_rules at the integration level

[mwojtyczka]

LGTM