Add standalone tests for LLM behavior regression testing

[neuromechanist]

Summary

The CI workflow has a standalone-tests job that runs on PRs targeting main when LangGraph components are touched, but there are currently zero tests marked with @pytest.mark.standalone. This job needs actual tests to fulfill its purpose of verifying LLM behavior hasn't regressed.

Context

• The standalone marker is defined in pyproject.toml and intended for tests that "run locally without backend (requires OPENROUTER_API_KEY_FOR_TESTING)"
• The CI job was recently fixed to only trigger on PRs targeting main to avoid unnecessary API costs
• There are @pytest.mark.integration tests in test_integration_openrouter.py, but none with @pytest.mark.standalone

Requirements

• Create standalone tests that verify core LLM annotation quality (e.g., known event descriptions produce valid HED annotations)
• Tests should use OPENROUTER_API_KEY_FOR_TESTING and a cheap model
• Cover key regression scenarios: invalid tags not generated, proper grouping, semantic faithfulness
• Consider golden-reference test cases (known input -> expected valid output patterns)
• Keep tests minimal to control API costs; focus on grounding behavior, not exhaustive coverage
• All tests should be marked with @pytest.mark.standalone

Related

• PR Fix LSP tag feedback loop and Safari SSE streaming #101 fixed the standalone CI trigger to only run on PRs to main
• The standalone marker is already registered in pyproject.toml (line 128)