fix: return ERROR not CAUGHT when evaluator cannot perform comparison

[MdSadiqMd]

Closes #24

Summary

• Path-walk failures and other "evaluator can't compare" cases in eval_modes.py returned CAUGHT, inflating hallucination metrics and triggering heal-loop retries that can never converge
• Changed six sites to return ERROR — aligning with the evaluator's own docstring: "ERROR — the evaluator could not actually evaluate. This is not evidence of tampering and must not be conflated with CAUGHT"

Line	Situation	Comparison happened?	Before	After
36	Path walk fails (KeyError/IndexError/TypeError/ValueError)	No, value never extracted	CAUGHT	ERROR
46	Unknown verification_mode	No, evaluator doesn't know the mode	CAUGHT	ERROR
74	expected_json_schema is empty for schema_type	No, nothing to validate against	CAUGHT	ERROR
80	SchemaError — the schema itself is invalid	No, broken schema can't validate anything	CAUGHT	ERROR
86	range_min and range_max both None	No - no bounds to check against	CAUGHT	ERROR
90	Indexed value at path isn't numeric	No, can't perform numeric comparison	CAUGHT	ERROR

Test plan

• uv run pytest tests/unit/ -> 116 passed
• uv run pytest tests/e2e/ -> 15 passed
• Updated test_field_extraction_error_when_index_out_of_range to expect ERROR
• Updated test_schema_type_missing_path_is_error to expect ERROR
• Verified real-mismatch tests (test_evaluate_handoff_caught_on_mismatch, test_final_verify_failure_marks_caught, scenario B tampered claim) still return CAUGHT