[Yushu-Li]

Problem and change

Successful cascade stages can return numeric scores alongside boolean flags. The threshold fallback counts those flags as numbers, and merging stages converts them to floats. For example, {"score": 0.1, "valid": True} passes a 0.5 threshold and can acquire inflated fallback fitness after a merge. A false flag can also block an otherwise passing score.

I exclude boolean flags from fallback cascade thresholds and preserve their type when merging stages 1/2/3, so the existing fitness helper continues to ignore them. Numeric scores still normalize to floats. Numeric combined_score precedence, inclusive thresholds, error filtering, and timeout policy are unchanged. This is a separate successful-cascade path from the timeout work in #446.

Validation

• Seven new test methods exercise actual local cascade modules, stage advancement, final fallback fitness, boolean-only results, and numeric controls. The regression cases fail on the original implementation.
• New tests plus existing cascade validation: 18 passed.
• Full documented command OPENAI_API_KEY=test-key-for-unit-tests python -m unittest discover tests: 997 tests, OK.
• Black checks and git diff --check passed.

No paid model calls were made. This does not add general validation of invalid combined_score values or change the policy for failed later stages.

AI assistance: I used OpenAI Codex for implementation, tests, and automated review.