[Yushu-Li]

A fallback fitness made from finite metrics can become infinite because the intermediate sum overflows: averaging two 1e308 values currently returns inf, even though the correct average is 1e308. This can distort fitness comparison and prompt feedback.

I rescale the average only when its sum is non-finite but every input is finite. Ordinary averages keep their existing calculation; explicit infinity handling, boolean exclusion, feature filtering and combined_score precedence are unchanged.

Validation:

• Five new regression/control methods; the original implementation fails three methods (16 failing subcases), with two control methods passing.
• Full unittest discovery: 1032 tests run, 2 skipped, no failures.
• Changed-file Black and whitespace checks passed.
• No model/provider calls. This does not change the policy for explicit non-finite evaluator outputs or the sum utility.

I used Codex assistance for implementation and regression testing.