Simplify agent instructions and skill routing

[Noisemaker111]

Problem

Agent guidance is duplicated across root policy and oversized skills, so normal tasks ingest unrelated manuals and contradictory workflow ceremony.

Context

• README.md already owns stable repository facts, but CLAUDE.md, AGENTS.md, and domain skills repeat them.
• The jgengine router contains domain manuals instead of routing progressively.
• Legacy size baselines grandfather oversized intake, while API/orphan checks can reward symbol mentions rather than usable guidance.
• PR Un-red main: fix orphan-gate false positives, guard degraded generation, surface red main #945 changed API, capability, and orphan-gate data; this work must re-measure and preserve its useful protections.

Suggested scope

1. Keep CLAUDE.md and AGENTS.md identical, governance-only, and under an enforceable few-KB budget.
2. Reduce jgengine/SKILL.md to a true router of at most 100 lines; consolidate aliases and give each domain one owner.
3. Add active-intake, mirror-equality, routing, and duplication checks that prevent prose dumping without weakening safety.
4. Preserve editor ownership, upstream-first design, layering, prohibited assets, attribution, user-owned merge/release, and risk-based verification.
5. Prove lower root, total skill, router, and normal-game active-intake bytes; update generators, references, tests, and skill packaging.
6. Ship the coherent overhaul in one PR that documents the reviewable/revertible PR-boundary rule.

[Noisemaker111]

Working on this in codex/instruction-system-overhaul.