IrredenEngine: Skill for Claude Code

Render Verify

render-verify is the "did this change break anything I wasn't looking at?" gate; render-debug-loop is the iterative build → run → inspect loop for chasing a known bug, and attach-screenshots produces the PR-body before/after pair. scripts/render-compare.py is the comparator on its own for one-off diffs. Validator index: docs/agents/VALIDATION.md.

Prerequisites

• Any preset (docs/agents/BUILD.md). The harness reads the active backend from build/CMakeCache.txt and looks up references under creations/demos/<demo>/test/references/<preset>/. Backends produce pixel-different output, so each keeps its own reference set — never shared, and never committed from a host that cannot build that preset.
• Screenshots are framebuffer-sized: a Retina host captures at 2× and fails a non-HiDPI reference on shape mismatch. Capture references on a host representative of where verify runs.
• The demo implements --auto-screenshot [warmup-frames] (engine/video/CLAUDE.md §"Auto-screenshot helper") and is deterministic across runs — random or time-dependent output makes the harness meaningless.

Manifest

creations/demos/<demo>/test/references/manifest.json lists the shots in the demo's shot-table order (the Nth shot maps to the Nth numbered screenshot):

{
  "demo": "shape_debug",
  "target": "IRShapeDebug",
  "screenshot_subdir": "save_files/screenshots",
  "shots": ["zoom1_origin", "zoom2_origin", ...],
  "thresholds": {
    "per_pixel_tol": 8,
    "match_pct": 99.9,
    "max_delta": 64,
    "psnr_db": 35.0
  }
}

"target" is authoritative for --target → demo-directory resolution: the harness scans every manifest and matches on it, falling back to inferring the directory from the target name (strip IR, snake_case) only when no manifest declares it. A manifest without target is invisible to --all and to the lookup.