chore: v9.3.0 V10 Agentic Reform

[nullhack]

Summary

V10 agentic infrastructure reform addressing 7 failure modes observed in V9, plus simulation-driven fixes.

Changes (12 files)

AGENTS.md — Full Rewrite

• Caveman-style throughout (verb-first, no articles)
• Main Directive: 4-point recovery anchor after compression
• Golden Rules tightened with feature lifecycle phases
• Runtime artifact resolution convention
• Review loop relay protocol (findings survive re-dispatch)
• Loop exit detection in anchor
• Multi-skill ordering explicit (listed order = execution order)
• Beehave violations = test failures, auto-generation docs

Skills (4 files)

• refactor/SKILL.md: beehave check gate + loop exit detection via beehave status --json
• write-test/SKILL.md: no docstrings in tests, self-heal ruff
• review-gate/SKILL.md: enforce no docstrings in structure review
• structure-feature/SKILL.md: git_branch Runtime artifact clarification

Knowledge (4 files)

• todo-anchor-protocol.md: synced notation, loop resolution, review relay
• test-design.md: comment/docstring noise pattern
• test-stubs.md: auto-generation pipeline, stale stubs, beehave clean
• skill-design/principles.md: imperative language standard

Config

• pyproject.toml: "D" added to per-file-ignores for tests/**

Simulation-Driven Validation

Walked full 24-state define-flow → develop-flow path for a simple "Add Task" feature. Identified and resolved 5 critical pain points:

1. Runtime artifact ambiguity → resolution convention
2. Feature file lifecycle → three-phase documentation
3. Multi-skill ordering → explicit in dispatch
4. Review feedback survival → relay protocol
5. Loop exit detection → beehave status + anchor integration

Version

9.2.1 → 9.3.0