The finding worth keeping: this screen spans 10 domains across the two systems, so loading whole domain shards costs 76,480 tokens, while fetching exactly the 40 entries it uses costs 11,003 — 7.0× cheaper. Sharding is the right answer for a model building one domain and the wrong answer for a product spanning several. That argues for an MCP tool that returns one contract by id, not for a smarter way of slicing files. Both full indexes together are 153,502 tokens, which is the whole context window of most locally-run models.

What building this demo found: when it was first assembled, six of these forty surfaces came from institutional entries that carried no four-cell decision register — a demo arguing that the register is binding, with six components that had nothing binding to show. Design mode reported each as an explicit gap rather than hiding it, which is how they were noticed. Pulling that thread found 123 of 251 entries in the same state. All of them have since been resolved: 240 now carry a full register, and the remaining 11 carry a stated reason no use-or-not decision exists — a token set is what components decide with, not a component anyone chooses. The generator now refuses to build a catalogue entry that is neither. Six components on a demo page were the visible part of it.