scientific(split): reject evaluation identities already present at prior cutoff #695

Finding

admit_rolling_origin_partition(...) verifies that evaluation rows have an AvailableTime after the training cutoff in the evaluation snapshot, but it does not reject an evaluation identity that is already present in the training snapshot.

That allows a caller to omit an already-known identity from training_document_ids, rebuild the same UUID in the later snapshot with a later AvailableTime, and admit it as newly available held-out evidence. The selected train/evaluation ID sets are disjoint, so the current overlap check does not catch this cross-snapshot contradiction.

Scientific impact

Rolling-origin evaluation must not count evidence as future/new when the prior cutoff snapshot itself proves the same identity was already available. Otherwise temporal leakage can be introduced through caller-selected omission plus later availability rebinding.

This is a local split-integrity invariant, not source-provenance authentication: Evidence/source authority remains with #527/#658/#671. The split owner only needs to trust its own prior snapshot presence as proof that the identity is not newly available.

Required contract

• realistic RED: same UUID exists in the training snapshot at/before the train cutoff, but the caller excludes it from the training ID slice and supplies it in the evaluation snapshot with a later AvailableTime; admission must fail closed;
• preserve legitimate later training of documents that were genuinely absent from the prior snapshot;
• do not weaken connected-group, cutoff, duplicate, or availability checks;
• keep scientific(split): derive leakage-safe expanding rolling-origin history at the split owner #691 expanding-history behavior intact;
• document the distinction between prior-snapshot presence and external source authenticity.

Acceptance

admit_rolling_origin_partition(...) rejects every evaluation identity already contained in the bound training snapshot before numerical consumers can score it.