feat(consensus): pin Wan2.2 as genesis architecture

[vex0209-bt]

Summary

• pin leoma/wan2.2-i2v-a14b-diffusers as the genesis and base model at immutable Hippius digest sha256:79e4cc89edd935e8ea843ae2bd80fb85d3444128bb946990109a33ffafb2184e
• make the base-model digest part of consensus and validate both Wan2.2 transformer experts during architecture prescreen
• honor the consensus-pinned model-level CPU-offload policy required to leave safe H100 activation headroom
• pin the published leoma-source corpus manifest and harden its build/publish CLI validation
• freeze the CUDA evaluation dependency graph and protect every eval-server route with shared Bearer authentication
• add a four-process 8xH100 PM2 configuration, deployment examples, and a production rehearsal runbook

Why

The previous seed was not a usable immutable production genesis, and architecture prescreen referenced a mutable base with a synthetic revision. Wan2.2 A14B also has two transformer experts, so validating only the original transformer and placing the complete pipeline directly on one GPU would make prescreen incomplete and leave unsafe VRAM headroom.

This change makes the model, corpus, generation placement, and executable evaluation stack explicit and reproducible across validators.

Consensus impact

• consensus digest: sha256:9ce264ab7bbeff290d97bf15bd5ce06064d55c706d38319a44f6180de59f089f
• generation offload is pinned to model
• source corpus is pinned to manifests/leoma-testnet-v1.json at sha256:723a48e655ee7087bd7750892b8e98bc48bd7b61b5097de7e55a020c26a8c159
• validator and evaluation hosts must deploy the same commit before processing duels

Verification

• 628 passed
• uv lock --check
• git diff --check
• Docker Compose configuration validation
• clean installed-wheel import and CLI-entrypoint checks
• Wan2.2 config-only snapshot passes the dual-transformer architecture lock

Remaining production gate

No full-weight H100 duel was run in this workspace because it has no NVIDIA GPU. Before mainnet launch, operators must preload the approximately 126 GB snapshot on the 8xH100 host, verify the complete corpus, calibrate all eight physical H100s, and complete at least one real king-versus-challenger duel without OOM or watchdog failure.