HyperPlonk: Zeromorph partial reduction (canonical layout breaking change)

[kh0ra]

Why

mosaic-hyperplonk's claim-reduction step currently approximates the multi-linear extension opening. The proper Zeromorph (or PST / Gemini) reduction collapses the MLE evaluation into a univariate KZG opening that we verify via the existing alt_bn128 pairing syscall.

Sub-tasks

• Pick reduction strategy: Zeromorph vs PST vs Gemini. Zeromorph is closest to what Espresso ships.
• Implement the partial reduction step (canonical layout breaking change — bump major version)
• Update fixture format
• Re-run differential test against Espresso (#__)
• Re-measure CU budget; expect significant change

Blocks

Real HyperPlonk soundness. Currently Phase-3 scaffold only.