test-campaign

{
  "name": "test-campaign",
  "version": "0.20.1",
  "description": "Run a complete UI test campaign against an application and leave behind a living evidence page \u2014 coverage, requirements, user-flow storyboards, screenshots, component atlas and defects in one browsable surface where every row has a stable id somebody can point at. Reads the project first \u2014 Overview, PRD, feature specs, design md and the latest mock UIs \u2014 so the campaign knows what the product *claims* to do before it looks at what it renders, then enumerates the correctness space (surface \u00d7 state \u00d7 viewport \u00d7 theme \u00d7 role \u00d7 locale \u00d7 data shape \u00d7 modality \u00d7 execution plane \u00d7 oracle), samples it deliberately and says so, writes and runs the suite in the project's own harness, sweeps for what no requirement named, and measures the build against its design of record on structure, style, vocabulary and geometry rather than on pixels. Every case carries which rung of oracle it stands on, so a critical flow proved only by \"the element exists\" fails the gate instead of passing quietly; a case claiming pixels must name a real capture and the channel it came from; a lane claiming the app was running and drawn must name the artifact and what witnessed it attaching, because a suite once reported 100% checked over two desktop apps that had never drawn a window; a published screenshot must name what the capture channel was pointed at, because a wall of 20 captures once showed three unrelated documents while every gate passed and only the filename bound a picture to its surface; a surface's controls and a navigation shell's destinations are counted the way surfaces are, because a campaign once reported 32 of 32 cases passing and armed over an application whose six menu items opened one screen and whose every button ran an empty closure, and an effect is never the product's own report that it acted; a check the instrument could not perform is inconclus