[P0] Separate obedience tests from policy-bypass tests

[Petsku01]

Problem

Many attacks test 'echo this token' with substring detector = tautological. Proves obedience, not vulnerability.

Tasks

• Tag attacks as obedience vs policy-bypass in attack metadata
• Add attack_type field to schema
• Report obedience and policy-bypass rates separately
• Update METHODOLOGY.md to explain the distinction

Acceptance Criteria

• Results clearly distinguish 'model followed instruction' from 'model violated safety policy'