feat(eval): consume sharded corpus-v2 exams

[vex0209-bt]

Summary

• add version-gated loading for corpus manifest v1 and sharded v2 manifests
• fetch only shards required by the deterministic exam sample selection
• verify root, shard, MP4, PNG, decoded-video, and first-frame integrity before evaluation
• use the corpus caption and verified PNG first frame as TI2V conditioning inputs
• extend corpus verification, audit records, calibration support, tests, and the production runbook

Safety

• hash root manifests and shards before parsing
• fail closed on schema, content-address, decoding, or conditioning-frame mismatches
• preserve v1 compatibility and leave the active chain.toml manifest pin unchanged
• require no runtime captioning model, leaving all eight H100 GPUs available for four evaluation pairs

Verification

• 672 passed in the full test suite
• adversarial corpus-v2 parser, shard-selection, and artifact-integrity coverage
• wheel build and isolated wheel import passed
• git diff --check and changed-file static checks passed