How should AI roto/matting outputs be validated against production-grade mattes?

SAM 2/3, Cutie, MatAnyone 2 and CorridorKey produce masks or alphas differently (binary masks vs un-premultiplied foreground + linear alpha). What objective metrics and test plates (hair, motion blur, transparency, spill) predict whether an AI matte survives a real composite, and how do edge colour contamination and premultiplication conventions break handoffs between these tools and Nuke/Fusion/Resolve? Why this is still open: Each project reports its own benchmarks; no cross-tool evaluation on production plates was found.