Review Componentization

code-review reads the diff; visual-verify reads the pixels; this reads the conversion judgment — the class of error where the screen renders correctly and every gate is green, but the componentization is wrong: a block extracted with no evidence, a component named for an internal id, a screen re-telling states its regions already own. That judgment only survives review in a fresh context that never saw the conversion being made — the agent that just settled and extracted knows what each choice meant, which is exactly the knowledge that hides the defect.

Gate

Applies when a design study has just been componentized (a settled screen, an inventory, extracted components) or the user asks to review one. Assumes the mechanical gates are already green — lint, type-check, module boundaries, the design-token check, story tests. If they are not, run those first: this review spends its attention only on what tooling cannot decide, so a finding a tool already catches is out of scope here.

1. Assemble the judge's inputs — artifacts only