ui-craft v1.0.19

The landing recipe shipped three ASCII diagrams, one per content strategy, and the first was labelled default. Its geometry was text on one side and a visual on the other — which is the split composition class, the one the MCP's draw deliberately offers last because it is the fold every generator reaches for unprompted. Across seven measured builds, every fold that got built was one of the three geometries those diagrams prescribed, and none was the class the draw had offered. One build drew stacked, declared stacked, and built split: it followed the draw when asked what it intended and the recipe when it wrote the code, because the recipe is always in context, has a picture, and says default.

So the pictures are gone and the content grammar stayed. Step 1 is now two decisions in order — what the fold argues, which is yours to choose from the brief, and how it is laid out, which is drawn. No strategy is a default any more. Every numeric limit the diagrams carried is preserved, and three exclusions are stated where a geometry cannot hold an argument: a product-forward fold cannot be type-only, a message-forward one cannot be product-dominant, a proof-forward one cannot be a full-bleed overlay.

The skill also names fold_candidates and check_fold for the first time. It previously mentioned composition classes exactly once, in the brief's spent list, so the protocol existed only inside a tool's own response — an agent had to call the tool to learn it should call the tool. That is the likeliest reason three builds in seven never drew a class at all. Measured after the change, across three builds: the draw was called in all three, and verified against a declared class in two.

Installers and generated launchers now write ui-craft-mcp@0.8.4, which carries the matching half — an omitted comparison in check_fold no longer reads as agreement, and the draw asks for one committed fold rather than three nobody built.

Honest limit on all of the above: three builds is a direction, not a result. The convergence itself did not move — both verified builds still landed on band while declaring something else, and removing the competing diagrams did not stop it. Adoption turned out to be fixable by naming the protocol where the agent already reads. The geometry is not fixable by asking.