release(topic): bind topic-index coordinates to the fitted topic basis #663

Finding

Draft #639 now retains coordinate-bound diagonal-Laplace variances, but the released tepp.trsl_topic_lineage.v2 coordinates are still identified only by artifact-local integer topic indexes. ReferenceTopicModel retains the fitted topic_term_probabilities and document_topic_proportions, while TopicLineageArtifact projects neither a fitted topic-basis identity nor a reconstructable topic basis. As a result, sequence_edges[*].topic_index and diagonal_laplace_uncertainty[*].coordinates[*].{numerator_topic_index,reference_topic_index} are not bound to the fitted topic content that gives those coordinates scientific meaning.

This leaves two distinct release risks:

• two fits with different topic-term bases can in principle project the same currently retained counts/edges/variances/configuration coordinates and therefore lack an explicit mismatch signal for the omitted topic basis; and
• topic-model label permutation means a bare integer index is not a stable scientific identity across equivalent fitted-topic orderings.

This is not a request to infer semantic labels with an LLM. It is a deterministic fit-coordinate/release-identity problem.

Required RED / contract

Before changing the unreleased v2 projection, prove at least:

1. two otherwise compatible fitted results whose topic-term basis differs cannot produce the same released topic coordinate identity;
2. every sequence-edge topic coordinate and every ALR numerator/reference coordinate is bound to an owner-derived topic identity, not a caller-authored label or detached integer alone;
3. a topic-basis mismatch fails closed before a result is promoted/consumed;
4. the contract remains deterministic under ordinary input ordering that does not change the admitted fitted topic basis;
5. no semantic-topic naming, LLM interpretation, or stronger posterior claim is introduced.

The RED should exercise an actual fitted/reference-model projection where practical. Synthetic hand-authored rows may cover serialization edge cases but are not sufficient scientific acceptance by themselves.

Minimal owner direction

Prefer a content-bound opaque topic coordinate derived from the numerical owner output (for example a versioned digest over the exact fitted topic-term probability row and its admitted vocabulary coordinate), then project that identity alongside the current local index. Do not reorder topics post hoc merely to canonicalize labels: the current diagonal ALR variances are expressed relative to a fitted reference-topic basis, and arbitrary topic permutation would require a mathematically correct covariance/basis transformation that the standalone diagonal result does not retain.

If the topic identity requires vocabulary/source coordinates not yet immutably bound, keep this lane dependent on #658 Evidence/source provenance rather than hashing an unbound caller vocabulary and calling it authoritative. A digest alone is identity binding, not semantic reconstruction.

Scientific basis

The scientific concern is twofold. Finite mixture-family likelihoods are symmetric to component-label permutations unless an identification/alignment rule is supplied, so a local component number alone is not a durable scientific identity (Stephens, 2000). Separately, log-ratio coordinates on the simplex have meaning relative to a chosen coordinate/reference basis; changing that basis is a transformation of the represented coordinates rather than a string relabeling operation (Aitchison, 1982; Egozcue et al., 2003). These sources support the fail-closed basis-binding requirement; they do not define TEPP's wire format or authorize semantic topic labels.

APA 7th:

• Aitchison, J. (1982). The statistical analysis of compositional data. Journal of the Royal Statistical Society: Series B (Methodological), 44(2), 139–160. https://doi.org/10.1111/j.2517-6161.1982.tb01195.x
• Egozcue, J. J., Pawlowsky-Glahn, V., Mateu-Figueras, G., & Barceló-Vidal, C. (2003). Isometric logratio transformations for compositional data analysis. Mathematical Geology, 35(3), 279–300. https://doi.org/10.1023/A:1023818214614
• Stephens, M. (2000). Dealing with label switching in mixture models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 62(4), 795–809. https://doi.org/10.1111/1467-9868.00265

Boundary

topic_measurement owns numerical fitted quantities. Analysis owns the digest-bound consumer projection. #662's diagonal-Laplace quantity work remains valid and must not be rewritten as joint covariance. #658 remains the source-population/vocabulary provenance path, and #605 remains the Membership reconstruction path.

Keep #639 Draft. No source copy, mutable dependency, caller-authored topic labels, LLM topic naming, force push, destructive rebase, or gate weakening.

Refs #638 #639 #658 #662 ADR 0012.