v0.3.1: inference detector (4-tier), contradiction question-fallback, improved batch prompt (154 tests pass)

v0.3: fix all reviewed issues (bugs #1-3, issues #4-8, 125 tests pass)

v0.3: 2-call batch verifier, numeric consistency check, slot-aware relevance filter

v0.2: split verifier, span_matcher, calibration, frame-based contradictions