dsh-jev-interceptor: Jev Risk Classification and Auto-Approval Plugin for DeepSeek Harness

The `dsh-jev-interceptor` plugin integrates non-generative Jev models into DeepSeek Harness (DSH), replacing the existing high-cost generative LLM classification approach. It delivers low-cost, high-efficiency tool-call decision support by evaluating risk, irreversibility, and injection suspicion with millisecond-level latency. Core features include risk classification, evidence-based automatic approval, session reference takeover, shadow/forced mode switching, and a fail-closed mechanism. Controlled through commands such as `/jev-on`, the plugin prioritizes safety and automatically delegates to DSH’s default behavior in the event of an error, making it suitable for scenarios that require strict risk control and faster response times.