fdars 0.12.0

fdars 0.12.0

Milestone v13.0 — Scientific Provenance & Cross-Language Implementations. A code + docs + skill release: no fdars-core bump and no changes to existing binding behavior. Every fdars AI surface can now answer "what is method X, where does it come from, and how else could I do it?" with grounded scientific provenance.

Added — scientific provenance (new code)

• fdars_method_references(method) MCP tool on the existing server — returns author-verified foundational-paper provenance (title/authors/year/DOI/URL) plus cross-language (R / Python / Matlab) implementation pointers for a callable, or an explicit {"curated": false, "sentinel": "NO_CURATED_ENTRY"} for the uncurated tail. Provably LLM-free: static importlib.resources JSON load, frozenset-gated, AST-guarded against any advisor/provider import — no model or network call, no synthesis.
• python/fdars/_references_map.json — a committed, hand-authored, author-verified references map: 57 papers with a sub-method-keyed callable_index (243 entries) in the same identifier space as the capability map. Partial-but-honest coverage: 28 / 437 callables carry a curated primary-paper entry (437 = 409 public callables + 28 Fdata class methods); the rest carry full provenance flagged curated: false pending human DOI-landing-page verification. Anti-feature families are wired to the absence sentinel, never a forced citation.
• GATE-05 guard suite (tests/test_guard_sync_version_independent.py) — internal consistency, cross-file resolution, structural DOI/URL gate (offline; opt-in live resolve via FDARS_ONLINE_CHECKS=1), coverage-fraction emission, anti-feature-absence, an LLM-free-boundary AST check, and a three-way _REFERENCES_MODULES == _EXPECTED_REFERENCES_MODULES == _CAPABILITY_MODULES mirror.

Added — documentation & skill

• scripts/generate_capability_dataset.py --references — an OFFLINE emitter (no fdars import) that generates a family-grouped docs/references.md cross-index (method → paper → R/Python/Matlab implementation) and a ## Scientific Provenance & Cross-Language Implementations section in llms.txt, both with an honest derived coverage fraction. Page nav-wired and rendered under the whole-site mkdocs build --strict gate.
• fdars-capabilities skill — a ## Scientific Provenance Protocol: prefer curated data (return verbatim); on curated: false / sentinel a consumer MAY synthesize but MUST structurally flag it with a machine-readable grounded: false, never presenting synthesized provenance as curated. Non-duplicating the fdars-advisor boundary; both paths demonstrated and tested.

Notes

• Author-verification is a hard authoring rule — every curated: true entry was checked against its DOI landing page by a human (e.g. Fraiman–Muniz depth is 2001, not the widely-miscited 1991). Whole-site mkdocs build --strict green offline; grounding invariant + LLM-free MCP boundary preserved.