AI Workbench MCP

This is an acceptance gate layer for AI coding agents, built MCP-first for Goose with Codex local/IDE support coming. It gives you seven tools that open a run, select a validation profile, record agent output, run deterministic checks, apply a quality gate, and render evidence folders with task metadata, validation reports, and accept/needs-review/block decisions. The validation profiles run actual commands like tests and linters, check for required artifacts, and enforce changed-file policies. You'd reach for this when you want auditable proof that an agent run should merge, not just the agent saying "done". It also ships a PR gate workflow that reads the resulting evidence and blocks PRs missing acceptance artifacts.