Can AI models write Manim code from natural language prompts? I tested Gemini, GPT, Claude, and Kimi to find out what works, what breaks, and why spatial reasoning is the bottleneck.