Frontend Design Loop MCP

This is a screenshot-driven design improvement workflow for coding agents. It exposes frontend_design_loop_design to iterate on working-but-rough pages using vision-grounded feedback, frontend_design_loop_eval to validate patches with deterministic checks and screenshot proof, and a toolkit variant that hands planning and edit control back to the host agent. The iteration loop runs against a local preview server, captures rendered screenshots, and returns winning patches with visual artifacts. Reach for it when the base agent got the layout functional but the result still looks generic or flat. The main workflow keeps planning, generation, and vision scoring on one provider and model lane by default, making multi-model routing opt-in rather than required.