v2.1.0

CodeInspectus 2.1 adds two bounded AI-code security detections, improves React dangerous-HTML flow coverage, fixes a JavaScript scanner edge case, and hardens the release path with signed-tag verification and npm provenance.

Added

• ci-ai-llm-output-dynamic-execution follows recognized model output into JavaScript eval/Function, import-proven Node shell-string APIs, and import-proven Execa command-string APIs. The bounded intrafile rule includes a frozen 5-TP/7-safe corpus and explicit false-negative boundaries.
• ci-ai-nextjs-admin-route-no-authz checks conventional Pages Router and App Router admin API handlers for visible authentication plus a server-controlled role/permission decision. Supabase client-writable user_metadata is deliberately rejected as authorization evidence.
• The dangerous-HTML rule now follows one local destructured function-component prop hop into dangerouslySetInnerHTML, with sanitized, trusted, and text-rendering precision cases.
• A reproducible public-repository scan, fix, same-path rescan, and sealed-evidence workflow is documented in examples/reports/rich-v2.1.0-sealed-workflow.md.

Fixed

• The bounded JavaScript lexer now accepts valid regular-expression literals beginning with =, including /=.*/s, instead of misclassifying them as division assignment and making the full scan partial.
• CI now exercises the independent packed TypeScript SDK consumer and a native macOS ARM64 row in addition to the existing platforms.

Release inventory

• Package, MCP server, CLI, SDK API, MCP Registry manifest, and reference CI install versions are synchronized at 2.1.0; existing V2 export and storage schemas remain 2.0.0.
• Detection database 1.15.0 contains 88 curated detections: 67 first-party native rules across 16 packs, 18 Opengrep-owned SAST rules, and 3 Gitleaks rules.
• JavaScript/TypeScript pack 1.5.0: 10 analyzers and 24 rules.
• Aggregate native engine: codeinspectus-ai@5.15.0.

See the complete CHANGELOG.md and release process.