NS-0.1: baseline models for warehouse pick-and-place

This note summarizes our first public checkpoint on the internal manipulation stack. The goal is honest throughput and repeatability on real cells—not leaderboard scores on synthetic scenes.

NS-0.1 uses a two-stage pipeline: coarse grasp proposals from an overhead view, then closed-loop refinement with wrist-mounted depth. Latency budgets are documented in the appendix (placeholder).

We report mean cycle time and success rate over 10k picks across three customer-shaped fixtures. Figures will ship in a later revision.