[Feature] Export calibrated shadow comparison datasets with immutable provenance

[Xunzhuo]

Goal

Export calibrated primary-versus-shadow comparison datasets with immutable provenance for reviewed Router Model experiments.

Proposed Workgroup

Router Models & Inference Runtime

Scope

• Convert approved shadow observations into a versioned dataset manifest.
• Preserve prompt/reference identity, primary and shadow model identities, output hashes, judge or label provenance, sampling policy, and redaction state.
• Define deterministic inclusion, exclusion, deduplication, balancing, and train/eval split rules.
• Validate schema, lineage, privacy policy, and artifact immutability before export.
• Support pluggable storage destinations without committing datasets to the source repository.

Non-goals

• Capturing live shadow traffic; tracked by the shadow dispatch issue.
• Training or promoting a model.
• Treating an unreviewed judge score as ground truth.

Completion signal

• One approved shadow corpus exports reproducibly from the same inputs to the same manifest and split identities.
• Every example is traceable to immutable, privacy-reviewed evidence.
• Invalid lineage, redaction, or split leakage fails before publication.

Parent Epic: #2975