scientific(selection): freeze training prevalence transform for held-out evaluation #681

Finding

#680 requires leakage-safe rolling-origin candidate-K validation with a separately measured held-out/predictive diagnostic. The current topic_measurement::ReferenceTopicInput cannot yet project a later evaluation horizon onto the training prevalence coordinate system without leakage.

ReferenceTopicInput::new(...) calls build_design(...), and build_design(...) calls standardize_event_time(event_times). standardize_event_time computes its origin/mean/deviation from the event times supplied to that particular input. Those transformation parameters are not retained in ReferenceTopicInput or ReferenceTopicFit as an owner-issued prevalence-design basis.

A rolling-origin evaluator therefore has two unsafe choices today:

1. construct a separate evaluation ReferenceTopicInput, which standardizes EventTime using evaluation-horizon statistics before applying training prevalence coefficients; or
2. construct one combined input so train/evaluation rows share a transform, which makes the training design depend on future evaluation evidence.

Either choice violates the intended knowledge-cutoff boundary. The issue is not that future EventTime is intrinsically invalid (#671/#672); the issue is that the feature transform used by the fitted prevalence coefficients must be learned from the training state and reused unchanged for evaluation.

This also affects categorical prevalence features. build_design derives the ordered membership feature basis from memberships active in the input. A separately constructed evaluation input can therefore have a different feature basis/order from the training fit even when dimensions happen to be compatible.

Required causal RED / repair

Add an owner-issued, immutable prevalence-design basis retained by the admitted training input/fit and used for evaluation projection:

• retain the training EventTime standardization origin/location/scale required to reproduce training design rows exactly;
• retain the ordered PrevalenceFeature basis used by the fitted coefficient matrix;
• project evaluation rows onto that frozen basis without recomputing location/scale from evaluation-horizon rows;
• fail closed when an evaluation covariate/membership feature cannot be represented by the training basis rather than silently reordering/dropping it;
• prove A/B fixtures where the same evaluation observation is scored in two evaluation batches with different surrounding future observations: its projected design row must be identical because only the frozen training basis may control the transform;
• preserve owner-authenticated future EventTime that was already available by the relevant cutoff; do not reject it merely for being after the cutoff;
• do not copy Evidence receipts, Membership authority, or scientific(time): bind admitted document event times to cutoff-qualified source evidence #671 source/event-time authentication into this numerical basis.

After this prerequisite exists, #680 can add fixed-global-parameter held-out/document-completion scoring and rolling-origin recovery evidence without test-horizon preprocessing leakage.

Scientific boundary

This basis is numerical/prevalence coordinate authority only. It is not source provenance, relation promotion, Membership provenance, topic identity, or release activation authority.

Roberts, M. E., Stewart, B. M., & Tingley, D. (2019). stm: An R Package for Structural Topic Models. Journal of Statistical Software, 91(2), 1–40. https://doi.org/10.18637/jss.v091.i02

Refs #638 #639 #671 #672 #680.