feat: v4 Knowledge Activation architecture

[r007b34r]

Summary

Complete v4 evolution of the skill suite from "production-ready prompts" to a Knowledge Activation architecture based on Atomic Knowledge Units (AKUs).

What changed

• All 7 SKILL.md files rewritten with AKU 7-component schema: structured triggers: / boundaries: / continuations: in frontmatter, constraints-first layout, tool-call style procedures, progressive disclosure with <details> blocks
• Validators added (validators/ per skill): deterministic pre/post check scripts for governance-as-code
• Eval expansion: all skills now have 25+ trigger cases (10 should + 10 should-not + 5 near-miss + failure modes)
• CI automation: GitHub Actions workflow validates structure, forbidden APIs, token budget, eval coverage on every push
• Scripts: validate-all.sh, check-forbidden-apis.sh, run-trigger-evals.sh
• Docs: architecture.md (knowledge topology), design-evolution.md (v1→v4), README-zh.md (Chinese split)
• README simplified: English-only, concise, links to detailed docs
• License unified: Apache-2.0 everywhere (was MIT in individual skills)

Key architectural decisions

1. Constraints placed FIRST in every SKILL.md (attention decay resistance)
2. Knowledge Topology declared via continuations: — skills form a navigable graph
3. Boundary discipline machine-readable: owns / delegates_to / never_absorbs
4. Token budget enforced: all SKILL.md files 137-165 lines (well under 500 limit)

Validation

=== ALL CHECKS PASSED ===
All 7 skills: 25/25 eval cases ✓
All skills within token budget ✓
No forbidden API references ✓

Test plan

• Run bash scripts/validate-all.sh — all checks pass
• Run bash scripts/run-trigger-evals.sh — 25+ cases per skill
• Deploy to OpenCode and test inbox-triage on real vault
• Verify cross-skill handoff (triage → connection-review → note-promotion)
• Confirm CI runs on push

🤖 Generated with Claude Code