feat: Prism skill — foundation fix with research gate, decision boundary, operator boundary

[fraser-svg]

Summary

Prism is a creative founder's AI co-pilot — the layer between non-technical creators and gstack that makes them power users without needing to become one. This PR adds Prism to the gstack skill tree and fixes the foundation issues discovered during dogfooding.

What changed:

• Decision Boundary — single source of truth for what Prism decides silently (engineering) vs surfaces to the founder (product tradeoffs, cost, capability)
• Research & Grounding Gate — 60-second time-boxed research step before building chunks with external dependencies. Source hierarchy: official docs first, community examples second. Surfaces approach decisions as confident recommendations
• Operator Boundary — Prism never sends the founder to another terminal. If it can install, configure, or run something, it does it itself
• Code reviewer agent wired into chunk build cycle with degradation handling
• 6 contradictory "just build" instructions rewritten to reference the Decision Boundary while preserving momentum
• Precedence table expanded from 6 to 8 entries (research checkpoints + operator disclosures)
• Technical Decisions persistence in intent.md across sessions
• Behavioral E2E eval — X/Twitter crawler scenario that caught the original "watered-down Claude" regression

Why: First dogfooding revealed Prism skipped research (built a web scraper instead of checking the X API), made silent product decisions ("that would cost money"), suggested broken APIs, and felt like downgraded Claude instead of enhanced gstack.

Pre-Landing Review

• Claude review: 7 issues (1 critical — eng review doc contradiction, 6 informational)
• Codex review: 4 P1s (all pre-existing — git add -A safety, auto-push, sessions array, local-process-as-deploy)
• Codex adversarial: 9 findings including shell injection in logging, research gate bypass path
• All pre-existing issues saved as follow-up TODOs, not blocking this PR

Test plan

• bun test passes (exit 0)
• Behavioral E2E eval added (prism-x-crawler-judgment) — not yet run against live API
• Dogfood /prism on a real project to validate foundation fix

🤖 Generated with Claude Code