Fix: scorers crash on empty input (UnboundLocalError / ZeroDivisionError)

[jimdawdy-hub]

Problem

Two scorers raise on an empty generations list instead of returning a score:

• evaluate_successor_liability binds f1 only inside the per-sample loop → UnboundLocalError when generations == [].
• evaluate_definition_extraction ends in correct / total → ZeroDivisionError when total == 0.

An empty list is a normal degenerate case (a task whose generations all failed to parse, a zero-length slice, etc.).

Fix

Return 0.0 in both when there are no generations.

[jimdawdy-hub]

Reproduction (captured output)

# repro.py
from evaluation import evaluate_successor_liability, evaluate_definition_extraction
for name, fn in [("evaluate_successor_liability", evaluate_successor_liability),
                 ("evaluate_definition_extraction", evaluate_definition_extraction)]:
    try:
        print(f"{name}([], []) -> {fn([], [])}")
    except Exception as e:
        print(f"{name}([], []) -> {type(e).__name__}: {e}")

BEFORE (upstream main):

evaluate_successor_liability([], []) -> UnboundLocalError: cannot access local variable 'f1' where it is not associated with a value
evaluate_definition_extraction([], []) -> ZeroDivisionError: division by zero

AFTER (this branch):

evaluate_successor_liability([], []) -> 0.0
evaluate_definition_extraction([], []) -> 0.0

[jimdawdy-hub]

@neelguha Friendly ping on this one as well — another small scorer hardening fix.

What it fixes: Two scorers crash on empty generations instead of returning a score:

• evaluate_successor_liability → UnboundLocalError (F1 only bound inside the loop)
• evaluate_definition_extraction → ZeroDivisionError when total == 0

Scope: One commit; returns 0.0 for the empty-input degenerate case.

Relation to #50 / #51: Independent. Useful on its own for robust eval pipelines (failed parses, empty slices, etc.).

Let me know if you’d prefer these three successor_liability-related PRs squashed — happy to do that. Thanks!