Auto-snap images onto ai-toolkit training buckets (W×H ÷64, ≤1.64MP) via
cover-scale + center-crop (Lanczos), per KLEIN_BUCKET_SIZES.md. Pure stdlib
bucket math (reproduces the spec table) + a torch node that also transforms
an optional mask identically and outputs width/height/label. No frontend.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
