v5.5.0

Added — Karpathy editing-discipline guidelines (inline + skill)

Adopts the four behavioral principles from
forrestchang/andrej-karpathy-skills,
grounded in Andrej Karpathy's
observations on LLM coding pitfalls.
Two delivery surfaces:

1. Inlined in generated CLAUDE.md and AGENTS.md. New
   "Editing discipline (Karpathy-inspired)" section in
   templates/core/CLAUDE_MD_v2.md and
   templates/core/AGENTS_LEAN.md. Every project that runs
   rulebook init or rulebook update from 5.5.0 onward gets the
   four principles in both files; the AGENTS.md projection
   propagates to Cursor / Windsurf / Cline / Copilot / Codex /
   Gemini through their existing references.
2. New skill templates/skills/core/karpathy-guidelines/SKILL.md
   — a dedicated invocable surface (/karpathy-guidelines or
   "follow karpathy discipline"). Added to INVOCABLE_CORE_SKILLS
   so installSkillsFromSource copies it to .claude/skills/ on
   init/update. Skill body kept ≤80 lines; total context cost
   < 1KB per session.

The four principles:

1. Think before coding — surface assumptions, present
   interpretations, ask when unclear.
2. Simplicity first — minimum code, no speculative abstractions
   or unrequested flexibility.
3. Surgical changes — touch only what the request requires; no
   opportunistic refactors of adjacent code.
4. Goal-driven execution — define verifiable success criteria
   upfront so multi-step tasks loop independently.

Targets gaps not covered by existing Tier 1 prohibitions —
particularly surgical changes (no rule today forbids
opportunistic refactor) and simplicity first
(no-shortcuts.md forbids stubs but not bloat). Test counts
updated: agent-delegation.test.ts skill-dir count 19→20.