Use when judging the composited, assembled screen of a running Godot game — "does it look designed and read clearly?" — typically after the asset re-skin (before validator's styled gate) or standalone on any running game. Renders every game state, fans out one fresh auditor subagent per lens (inventory/completeness, fidelity/cohesion, composition/collision, legibility, colour-accessibility, structural-fidelity, polish/design-quality), dedupes + attributes findings, and drives the fix → re-render → re-audit loop. Records NOTHING to the manifest — outputs are code fixes (git) + a findings report.
