scientific(simulation): own true topic trajectories for realistic rolling-origin K recovery #689

Finding

#680 now has the numerical pieces needed to evaluate candidate K without temporal preprocessing leakage: #684 fixed-training prevalence-mean prediction, #685 owner-admitted newly available partitions, #686 exact partition↔numerical row binding, #687 one-window predictive K selection, and #688 contiguous multi-window score aggregation.

The remaining acceptance cannot honestly be built from the current tepp_simulation::TruthManifest. The simulation owner currently records latent events, documents, availability, memberships, true/observed relations, method variants and digests, but it does not record a true topic count, topic-term probabilities, document logistic-normal coordinates, prevalence coefficients/covariance, topic trajectories, or generated term counts. SimulationConfig likewise has no topic-measurement DGP parameters. Consequently an apparent rolling-origin K-recovery test would have to hand-author counts outside the truth owner or infer "truth" from the estimator input, defeating the TEST_STRATEGY requirement for true-parameter recovery.

Required owner work

Extend tepp_simulation rather than adding a model-selection-local fixture. The owner contract must be deterministic/reproducible and expose enough truth to build realistic rolling-origin recovery without source rewriting:

• explicit true K and vocabulary size with bounded nondegenerate topic-term probability matrices;
• declared prevalence trajectory parameters over EventTime, including stable/interpretable temporal drift and covariance rather than independent per-window topic identities;
• document latent logistic-normal coordinates and resulting topic mixtures generated from those declared parameters;
• weighted multilevel/cross-classified/multiple-membership contribution truth aligned with the generated document memberships;
• relation structure that can affect generated topic coordinates while retaining the existing event/revision/translation/copy truth and connected leakage groups;
• document term counts generated from the declared topic/content state with nonzero realistic document lengths and overlap/sparsity/imbalance controls;
• availability delays independent from EventTime semantics so deliberately future-available evidence can be constructed and rejected by corpus_split;
• immutable truth-manifest digest coverage for all topic DGP parameters, latent document states and term-count observations;
• deterministic CPU f64 generation with explicit seed domains and no LLM-generated numerical truth.

The public recovery fixture must not be a four-row micro-corpus. It should exercise multiple canonical rolling-origin windows, changing prevalence, overlapping topics, relation-connected observations and weighted memberships; candidate selection must consume #688 and each replication outcome must feed #680's denominator-preserving SelectedKRecoverySummary.

Required RED / acceptance

Before implementation, add a truth-owner RED demonstrating that a generated scenario exposes declared true_k, true topic/content/prevalence state and term counts whose identities align one-to-one with the manifest documents. The manifest digest must change when any scientific truth parameter or generated count changes.

Then add the #680 recovery acceptance separately: recover known K and declared trajectories/states over repeated realistic scenarios, report bias/RMSE/failure denominator/MCSE, and include a deliberately future-available document whose attempted admission is rejected/audited. Synthetic micro-data remains unit coverage; do not promote the owner generator itself as release acceptance without the full recovery experiment.

Boundary

tepp_simulation owns the known-truth DGP. topic_measurement owns the estimator. corpus_split owns rolling-origin/leakage admission. model_selection owns predictive K comparison/recovery summaries. Evidence, Membership and relation promotion remain #527/#658/#671, #605 and #675 respectively. Do not copy those owner contracts or treat simulation truth as production evidence authority.

Refs #639 #680 #688.