Ship native Codex plugin with Earned Enforcement

[SignalLayerLabs]

What changed

• ships marginal@marginal as a native Codex plugin with one-command install and clean removal
• adds privacy-safe hook normalization, authenticated local transport, conservative structured outcomes, and fail-open behavior
• introduces repository-scoped Earned Enforcement receipts, explicit candidate review, automatic drift demotion, and fresh evidence windows after safety failures
• adds isolated Codex 0.147.0 install/lifecycle/privacy/removal acceptance evidence
• promotes the package and documentation surface to v0.3.0, with benchmark results and install commands first in README and Pages
• includes the external Codex directory submission packet without claiming approval before OpenAI review

Why

MARGINAL should govern agent compute only when it can prove that its intervention is safe and cheaper than the work it avoids. Global installation therefore remains Shadow-only; repository Tool Enforcement must be earned from local, reviewed evidence and is revoked on drift.

Validation

• 439 passed
• Ruff check and format check
• strict mypy over src/marginal
• reproducible plugin-runtime build check
• official plugin and skill validators
• isolated native Codex 0.147.0 marketplace install, four-hook lifecycle, 100% exercised hook coverage, zero secret-marker persistence, and remove
• Python sdist/wheel build plus Twine validation