Proposal: Belt ships new features as agent-executed skills, not templates

[rakeshpetit]

For new features, Belt stops scaffolding code itself and instead emits a versioned, thoughtbot-authored feature skill that the user's coding agent executes. Belt stays deterministic and offline: no LLM calls, no agent invocation. Existing features (create, add notifications) are untouched.

Full write-up: docs/the-future-of-belt.md. Ten decisions recorded as ADRs in docs/adr/0001–0010, with the reasoning for each.

This PR is the proposal + decision record only — no implementation yet. If it lands, the build order is already broken into issues:

• belt agent namespace: generic feature-skill emit machinery #80
• State-management feature skill: Zustand variant #81
• State-management: Redux Toolkit variant + emit-time variant prompt #82
• QA script formalizing the manual feature-skill workflow #83