SC-NeuroCore v3.16.0
Minor release accumulating the hardening of 237 commits since v3.15.35.
Headline — provenance & citation integrity
Every neuron-descriptor DOI is now verified against its registry (Crossref, or DataCite for arXiv preprints). This release corrects the citations that had drifted or been fabricated in generated descriptors — a non-existent "Kilinc & Bhatt (2023)" (the model is in fact the Nagumo-Sato / Aihara sigmoid map), a fabricated "Jahns et al. (2025)", a recurring phantom "Bhatt" co-author across several cerebellar models, an arXiv identifier that pointed at an unrelated paper, and several digit-transposed or wrong DOIs. A new offline gate (tools/provenance/verify_dois.py + a committed DOI ledger checked by tests/test_provenance_doi_integrity.py) fails CI if any descriptor DOI does not resolve at its registry, so this class of fabrication cannot recur silently. Descriptor citeable coverage rose from 119 to 131 models.
Also in this release
- Verification safety-screen hardening (
CodeSafetyVerifierrejects AST-visible filesystem / process / network / dynamic-execution escape routes). - Precision-solver and precision-config width/budget hardening; MLIR emitter and IR type-checker docstring + coverage ratchets.
- Go acceleration and services namespace docstring ratchets.
- Auto-model→RTL numerics fixes taking the cosim-validated model set to 12 (including Hodgkin-Huxley and Connor-Stevens at Q16.16).
- Dependency bumps (fastapi, ruff, GitHub Actions group) and a green CI baseline: the OpenSSF Scorecard and Security Scanner lanes are both passing again.
Full detail: see CHANGELOG.md.