[PaulHax]

Comparison membership needs explicit scientific identities

A recipe or run name cannot identify a replicate, establish a retry relationship or select comparable evidence.

Save validated records and explicit selections

Add versioned protocols, attempts, artifacts and evaluations with relationship checks, sourced unknowns, budgets and metric definitions. Collation selects named attempts and evaluations and retains included, excluded and missing cells in an immutable content-identified snapshot. Read-only frozen-defender and co-training adapters preserve unresolved historical membership; study-specific rules establish comparability.

The frozen analyzer verifies the saved configuration and analyzer/helper identity and retains its source with the report.

Part of #29, #31 and #33. Based on #34.

[PaulHax]

The review dataset exercises the adapters with a real historical co-training SQLite store and retained frozen-Blue evaluations.

Both co-training executions remain distinct, with unknown trial and retry relationships. The frozen-study inventory records 9 expected trial cells, 0 missing cells and 4 unresolved exclusions. Included attempts and evaluations are not counted as additional replicates.

The archive includes the historical store, raw evaluations and verification commands. Its frozen-study inventory is derived; excluded original source archives prevent replaying the full original source-validated collation.