visual-passlisted

# Running the manual visual pass without a human **This skill runs as a forked Sonnet subagent** (`context: fork`). It sees none of the caller's conversation, only the arguments: the worktree path, the quickstart section or the change to check, and what counts as a pass. It returns one verdict per check with the paths of the cropped screenshots that show it, so the caller can open a crop when a verdict looks doubtful. Screenshots stay in this subagent's context, not the caller's. Spec-kit features here end with tasks nobody can automate: *"§B1 and §B2 are this feature's two headline claims and neither can be automated; a green suite is not this feature working."* Those tasks sat unrun for four features because the only recorded route was a person at a display. There is another route. This skill is the verified one. ## What this catches that the test suite cannot The gates in `tests/` resolve **layout**. They compare rectangles, and they are exact about positions. They are structurally blind to: - colour, tone, and elevation (`style_snapshot` records values, not what they look like composited) - type weight — two labels at the same box with different weights pass every geometry check - **glyphs drawn on top of each other** — each node is where its own layout says it is - anything in `draw` only: `scale` and `fade` transform drawing, so an animating widget occupies exactly the boxes it occupies at rest The first real finding from this skill was a leading search icon dr