scientific(selection): require identical training state across recovery candidates #700

Finding

#693 requires every rolling-origin recovery window to cover the full predeclared candidate-K grid, and #694 binds each surviving fit to the exact candidate-specific numerical configuration. The recovery path still permits one scientific comparison defect: two candidate K fits can use the same training document identities but different training numerical state (term counts, EventTime rows, prevalence design, Membership weights, or admitted relation pairs).

rolling_origin_prevalence_mean_predictive_log_likelihood(...) currently binds a fit to a RollingOriginPartition by training identity set. RollingOriginRecoveryEvaluation::new(...) retains each fit's topic dimension/configuration but does not require all candidates in a window to have been fit from one exact ReferenceTopicTrainingInput. A K=2 fit trained on corpus A and a K=3 fit trained on dimension-compatible corpus B can therefore enter the same recovery comparison when A/B use the same document UUIDs.

For scientific K recovery this is invalid. Candidate K is the experimental factor; the training observations and frozen prevalence coordinates must be held constant within a replication/window. Otherwise score differences are confounded with different training data and can make true-K recovery non-reproducible.

Required RED / repair

• Add a realistic contract using owner-issued ReferenceTopicTrainingFit values with the same training document IDs and declared scientific(selection): bind recovery candidate fits to declared numerical configuration #694 configuration, but alter at least one training count row for one candidate.
• The recovery composition must fail closed before predictive scoring rather than accepting the candidates as comparable.
• Keep the comparison inside topic_measurement owner semantics: expose only a numerical training-state equivalence contract sufficient to compare exact retained training inputs; do not mint Evidence/source provenance authority or copy source receipts.
• Preserve scientific(selection): aggregate candidate K across canonical rolling-origin windows #688 generic operational survivor semantics. This stricter invariant belongs to the scientific-recovery path.
• Do not compare only document IDs or only PrevalenceDesignBasis: term/count, event-time/design, Membership/relation-derived design and transition coordinates must be covered by the retained owner input.

Boundary

This is numerical experiment integrity only. #527/#658/#671 remain source/Evidence/event-time authentication authorities, #605 remains Membership authority, and #675 remains observed-relation activation authority. A matching training state does not prove external provenance.

Refs #639 #680 #681 #683 #693 #694.