Make Napkin explicit and evidence-driven

[justin808]

Summary

• replace Napkin's unconditional session-start mutation with explicit lesson capture
• route mandatory rules to AGENTS.md, operational guidance to existing runbooks, and fallback lessons to one file per lesson
• disable implicit invocation in both Codex and Claude Code metadata
• document separate $napkin and /napkin invocation syntax

Why

Always-on curation crossed read-only boundaries, duplicated native memory and repository guidance, and added prompt ceremony to every task. The new workflow preserves durable, evidence-backed lessons only when requested.

Validation

• parsed SKILL.md and agents/openai.yaml metadata and asserted both explicit-only controls
• confirmed ordinary Codex prompt assembly omits Napkin
• forward-tested a fresh agent against a disposable repository; it updated the existing AGENTS.md once and did not create the fallback lesson directory
• ran git diff --check
• ran codex review --uncommitted; final review reported no actionable findings

[justin808]

@blader This is ready for review and merge at HEAD ad41b3f2da9084ffb1565b8d0fa3ee687c4e0dc8.

• Type: agent-skill behavior and documentation
• Validation: metadata policy assertions and git diff --check pass; fresh-agent forward test routed the lesson to the existing AGENTS.md without creating fallback memory
• Review threads: none
• Merge state: clean and conflict-free
• CI: this repository has no configured workflows or reported checks
• Known risk: explicit invocation replaces the previous always-on behavior by design

The PR author cannot merge or formally request reviewers because the fork account has read-only access to the upstream repository.