refactor(zhiyi): v0.3 architecture rebuild — progressive disclosure + provenance + fidelity tests

Sample master (zhiyi) rebuilt to new architecture:
- SKILL.md slimmed from 225→94 lines with decision tree + Quick Ref
- Provenance frontmatter: CBETA IDs, FoJin text IDs, citation_format
- voice.md/teaching.md moved to references/ (loaded on demand)
- sources/ with canonical excerpts (offline-capable)
- tests/fidelity.jsonl: 5 Q&A pairs with expected citations/keywords
- scripts/validate.py: cross-master frontmatter linter
- scripts/test-fidelity.py: Claude API-based fidelity test runner

Follows Anthropic Agent Skills patterns:
  - Progressive disclosure (metadata→body→references)
  - Decision tree for branching workflows
  - Task-gated reference loading
  - Quick Reference table
  - Scripts as black boxes (--help, never Read source)

prebuilt/zhiyi/tests/fidelity.jsonl

@@ -0,0 +1,5 @@
+{"q": "什么是一念三千？", "must_cite": ["T1911", "摩訶止觀"], "must_mention": ["十法界", "三千", "一念心"], "difficulty": "basic"}
+{"q": "三谛圆融是什么意思？", "must_cite": ["T1716", "法華玄義"], "must_mention": ["空", "假", "中", "圆融"], "difficulty": "basic"}
+{"q": "五时八教怎么分？", "must_cite": ["T1716"], "must_mention": ["华严", "阿含", "方等", "般若", "法华"], "difficulty": "basic"}
+{"q": "止观怎么修？初学者应该从哪里入手？", "must_cite": ["T1911"], "must_mention": ["小止观", "调息", "二十五方便"], "difficulty": "intermediate"}
+{"q": "一心三观和次第三观有什么区别？", "must_cite": ["T1911", "摩訶止觀"], "must_mention": ["空假中", "一心", "次第", "圆教", "别教"], "difficulty": "advanced"}