Stop That Shit 0.1.1 — Pi and Evaluation Refinements

在 0.1.0 的四套宿主基础上加入 Pi，并把评估证据拆开。 0.1.1
增加 Pi 原生 Extension，同时补齐 routing、Hook decision、task
acceptance 和 host effect 的独立观测边界。

Adds Pi to the 0.1.0 host set and separates evaluation evidence.
Version 0.1.1 adds the native Pi Extension and keeps routing, Hook
decisions, task acceptance, and host effects as separate observations.

新增 / New

• Pi Adapter：支持 @earendil-works/pi-coding-agent 0.84.4，复用共享
  controller、Skill 和 metadata-only Runtime；文档化的 subagent 形式在
  父工具调用处做原子预算控制，不宣称跨进程继承契约。
  / Adds the Pi Adapter for @earendil-works/pi-coding-agent 0.84.4,
  reusing the shared controller, Skill, and metadata-only Runtime. Documented
  subagent forms are budgeted atomically at the parent call; cross-process
  contract inheritance is not claimed.
• Paired evaluation vNext：加入不可变 Skill/CaseBundle digest；将 22 个
  implicit-routing case 拆为 required、optional、irrelevant 三类，并单独验收
  行为；三单元 host integration smoke 覆盖 mode deny、file-lock deny 和 allow。
  默认成对矩阵为 16 cases × 3 arms × 3 runs = 144 个 session，默认仍只生成
  计划。
  / Adds immutable Skill and CaseBundle digests. The 22 implicit-routing cases
  separate required, optional, and irrelevant routing from behavior acceptance.
  A three-cell host integration smoke covers mode deny, file-lock deny, and
  allow. The default paired matrix is 16 cases × 3 arms × 3 runs = 144 sessions
  and remains dry-run by default.
• 发布与验证边界：保留 Good Case、基础设施错误、Hook decision 与
  hostEffect 的分离，不把 deny 返回或离线 plan 包装成宿主阻断或模型效果
  结论。
  / Keeps Good Cases, infrastructure errors, Hook decisions, and
  hostEffect separate; a returned deny or offline plan is not presented as
  host enforcement or model-effect evidence.