AI startups don’t fail from bad models. They fail when decisions harden too fast, especially after Series A. A systems view of premature architectural lock-in.