scientific(selection): score only owner-issued reference fits #677

Finding

#676 now binds a statistical candidate's declared K to the fitted topic dimension, but model_selection::statistical_candidate_from_fit(input, candidate_k, model) still accepts ReferenceTopicInput and ReferenceTopicModel as detached public arguments.

ReferenceTopicFit was introduced in #665 specifically so an admitted input, deterministic configuration, and fitted model cannot be rebound after fitting. Model selection currently bypasses that nominal owner boundary: any dimension-compatible model can be scored against a different input, producing a finite in-sample likelihood and Schwarz-style score that looks statistically authoritative even though the model was not fitted to that corpus/design/membership/relation state.

The internal select_fitted_candidate_k(...) loop happens to pass the model it just fitted from the same input, but the public scorer remains a substitution path and the selector still uses the detached fit_reference_topic_model(...) result rather than the owner-issued aggregate.

Required RED

• construct two dimension-compatible admitted inputs A/B whose term evidence differs;
• fit or otherwise obtain an owner-issued reference fit for A;
• scoring the owner-issued A fit succeeds;
• substituting input B for the A model is impossible through the supported public scoring API or fails closed before emitting ModelCandidate;
• candidate K is derived/bound to the fit, preserving scientific(selection): bind candidate K to fitted topic dimension #676;
• select_fitted_candidate_k(...) uses the owner-issued fit path and retains existing fitted-selection recovery/LLM-refusal behavior.

Minimal causal direction

Make the public statistical scoring boundary consume ReferenceTopicFit rather than detached ReferenceTopicInput + ReferenceTopicModel. Migrate select_fitted_candidate_k(...) to ReferenceTopicFit::fit(...) and derive the candidate dimension from the retained model/config. Keep any detached likelihood arithmetic owner-internal or clearly non-authoritative; do not replace the owner aggregate with a locally computed digest.

If compatibility requires a staged API transition, the old detached scorer must not remain an equally authoritative public path indefinitely; record an explicit removal condition and keep CI warning-free.

Boundary

This is fit-local model-selection integrity. It does not prove Evidence source/vocabulary/event-time provenance (#527/#658/#671), relation promotion authority (#675), Membership authority (#605), fitted semantic topic identity (#663), or release projection (#664).

Refs #638 #639 #665 #670 #676.