bug: is_content_safe() drops policy categories for "safe" verdicts with trailing category tokens

Did you check docs and existing issues?

• I have read all the NeMo-Guardrails docs
• I have updated the package to the latest version before submitting this issue
• I have used the develop branch
• I have searched the existing issues of NeMo-Guardrails

Python version

N/A (logic bug, not environment-specific)

Operating system/version

N/A

NeMo-Guardrails version

develop (as of PR #2151)

Describe the bug

is_content_safe() in nemoguardrails/llm/output_parsers.py only extracts trailing policy-category tokens (e.g., S1, S8) when the response prefix is "unsafe". The "safe" branch unconditionally returns [True] and discards any trailing category text, so a response like "safe S1 S8" loses the category information entirely.

This was surfaced while asserting content_safety_check_input's RailOutcome.metadata["policy_violations"] in tests/test_content_safety_integration.py::test_content_safety_input_with_is_content_safe_parser_safe_with_violations, where the expected assertion result.metadata["policy_violations"] == ["S1", "S8"] had to be commented out because it can never pass with the current parser logic.

Steps To Reproduce

1. Call is_content_safe("safe S1 S8").
2. Observe the result is [True] instead of [True, "S1", "S8"].
3. Run content_safety_check_input(...) with a mocked task manager returning this parsed result and inspect result.metadata["policy_violations"].

Expected Behavior

When a response is "safe" but still lists categories (an edge case, but currently unhandled), is_content_safe() should capture and return those categories rather than discarding them, so downstream consumers like content_safety_check_input/content_safety_check_output can populate policy_violations metadata correctly.

Actual Behavior

Trailing category tokens after a "safe" prefix are silently dropped; policy_violations metadata ends up empty ([]) even when categories were present in the raw LLM response.

Affected files/modules

• nemoguardrails/llm/output_parsers.py (is_content_safe)
• tests/test_content_safety_integration.py (commented-out assertion in test_content_safety_input_with_is_content_safe_parser_safe_with_violations)

References

• PR: https://github.com/NVIDIA-NeMo/Guardrails/pull/2151
• Discussion: https://github.com/NVIDIA-NeMo/Guardrails/pull/2151#discussion_r3585336536
• Requested by: @Pouyanpi

Acceptance criteria

• is_content_safe() parses trailing category tokens regardless of the "safe"/"unsafe" prefix, or documents/handles the "safe with violations" case explicitly.
• The commented-out assertion in test_content_safety_input_with_is_content_safe_parser_safe_with_violations can be re-enabled and passes.