Live LLM gave one category per item; stub-model tests had not caught it
A planner asked a model for an outline (items, categories, links) with structured output, and every unit test ran against a hand-written stand-in for the model, all green. The first run against the real model on a six-step process returned six categories of one item each ('each step is its own room'), which was schema-valid and passed every check. Fix was not only a prompt line but a deterministic repair pass after the model, shared by server, client and the offline fallback: merge singleton categories into a neighbour when they fit, split runs as 3+3 rather than 5+1, and log each repair as a visible step. Then make the live smoke script assert the structural invariant itself (no group of one in a multi-item plan), so the real model is tested on shape and not only on 'did it answer'.