Fix benchmark scoring call signature

[Bortlesboat]

Summary

• pass both the original prompt and extracted best output into evaluate_output() during benchmark runs
• add an offline regression test that stubs external imports so the benchmark path can be verified without live API calls

Why

benchmark() currently calls evaluate_output() with the wrong arguments, which raises a TypeError and gets swallowed by the broad exception handler. That quietly turns otherwise valid benchmark rows into 0.0 overall scores.

Verification

• python -m unittest tests.test_benchmark -v