test: add property-based tests for evaluators using Hypothesis

[zurfjereluhmie]

Summary

• Add 21 property-based tests covering all 7 evaluators (numeric_tolerance, top_k, json_match, regex, exact_match, contains, membership) using Hypothesis
• Add hypothesis as a dev dependency
• Guard tests against _parse_expected's ast.literal_eval behavior that converts numeric strings (e.g. "0") into non-string Python objects

Evaluator coverage

Evaluator	Properties tested
numeric_tolerance	identity, bool return type, within-default-tolerance, non-numeric → False, formula verification
top_k	score in [0,1], perfect score, zero score, None output
json_match	score in [0,1], identical JSON = 1.0, invalid JSON = 0.0
regex	no crash on arbitrary strings, escaped literal matches itself, None output
exact_match	identity, different strings don't match
contains	substring detection, None output
membership	member found, non-member not found

Notable finding

_parse_expected uses ast.literal_eval on string expected values, which silently converts numeric-looking strings ("0", "1.5", "True") into their Python literal types. This causes type mismatches in evaluators that expect string inputs. Tests use assume() guards and a _survives_literal_eval helper to filter these inputs. This is a pre-existing design issue worth addressing separately.