Python parity of typescript/examples/05-evaluators. Exercises every generic evaluator directly:

output (+ target) -> Evaluator -> { status, score?, error?, messages?, evidence? }

• threshold_evaluator — pass by >= (default) or <= (at_most).
• numeric_error_evaluator — the output IS the error; score = 1 / (1 + |error|).
• exact_match_evaluator — structural equality vs the target.
• composite_evaluator — worst status wins, scores averaged, messages concatenated.

Evaluators are normally attached to a ModelGraph, but they are plain objects you can call on their own — which is what makes runs inspectable.

python3 python/examples/05-evaluators/main.py

Captured output: expected-output.txt.