v0.9.0 — Shadow Release

🦑 Safety-first guardian layer for LLM code fixing

Highlights

• OpenCode integration via EngineRouter
• Safety VETO with Base64/HTML decode
• SkepticBrain adversarial review
• Human Approval Gate
• BLM learning loop
• EngineLearning trust calibration
• Unified Bug Classifier (18 types)
• Two-Tentacle architecture

No auto-apply. No auto-PR. Human approval required.