Add `study_type` closed facet at Study level (primary vs secondary literature)*

[adelavega]

Add a closed study_type enum at the Study node (e.g. primary, meta_analysis, systematic_review, narrative_review, case_report, methods_development, simulation, commentary). This is the corpus-admission gate: keep primary, quarantine the rest.

Why this is an explicit gate, not a deferred-map facet

A meta-analysis or review paper's text is full of primary-study coordinates, contrasts, and tasks quoted from included studies. An open extractor will faithfully pull these as analyses — nothing about the extraction marks them as second-hand. If document type is left to emerge from a lower-level field (e.g. design_type), the failure mode is phantom training pairs: other studies' coordinates entering the corpus as if this paper produced them — the contamination class a contrastive objective learns as signal.

study_type belongs to the "decides whether the extraction should exist at all" class, so unlike descriptive facets it must be defined closed now and scored on its own with confidence, not deferred to second-pass mapping. Bonus: "classify this document's type" is a more natural, higher-accuracy LLM task than "should I exclude this."

Keep it distinct from two orthogonal axes it will be confused with: Study.design (observational/interventional, cross-sectional/longitudinal) and ONVOC's Study Design → Study Types (case-control, cohort, RCT). Both are the epidemiological-design axis; neither captures primary-vs-secondary literature type, which is a new axis we define ourselves.

See also Issue #7 (analysis-level data_origin) for the mixed-paper case this study-level gate does not cover.

[adelavega]

@jdkent commented this may be less important if we just use pubmed to gate out these papers

[jdkent]

pubmed is not perfect at annotating all the papers, but at least for meta-analyses, the word "meta-analysis" typically appears in the title as well, we can test to see how many meta-analyses this heuristic misses.