Design review

Input: application URL, setup recipe, observation/correction scope, authoritative decisions, and whatever reference exists: Figma frames, ticket mockups, .claude/tasks/design-reference.md, viewports, locales, consumer routes. When .claude/tasks/review-guide.md exists, read it and apply review guide. It is a task file, not product source. Return measured comparisons, findings that each cite their reference, coverage, unverified cells and reproducible observations for QA. The caller owns report formats and final verdict policy.

1. Pick the reference level with reference-levels.md and declare it. For Figma sources read the first part of read-design.md, "Reading a Figma design".
2. Build your own inventory from that reference and the brief, and write it down where the caller says, before reading the developer's measurements, captures or conclusions. A caller's inventory or style list never feeds or bounds yours.
3. Read a supplied setup recipe to reach the state, but inspect its simulation assumptions. Do not infer application internals from a DOM proxy or treat a malformed stub as a product bug.
4. Measure in the order and with the rules of visual-comparison.md, then run the checks of objective-checks.md: layout invariants, states, interaction design, accessibility, themes, wording, consumer routes. These need no design reference and apply at every level.
5. Observe the relevant screen and neighbouring alignment; classify unrelated existing differences separately so they do not expand the correction scope. Apply explicit decisions and source precedence supplied by the caller; if unresolved, report the conflict.
6. Reconcile with developer evidence only now, and return the differences between your inventory and what the developer measured.
7. Return expected/actual values, the reference of each finding, paired captures, limits and the coverage reached per screen, viewport and state. A missing source or unreachable state is unverified with its obstacle, never PASS.

No product source reading, code edits, service startup, user questioning or publication. If URL, credentials or reference access are missing, return the obstacle to the caller. Sharing measurement primitives or setup does not mean sharing the author's verdict.