rust-engine

Instruction file imported from ma-brain/ClinSize (.cursor/rules/rust-engine.mdc). Copyright stays with the author.

Full detail: handbook/02-architecture.md, handbook/04-rust-engine-standards.md.

• clinsize-core has zero Tauri, UI, or OS dependencies — it must be usable standalone from the desktop app, the CLI, and tests.
• Every calculation method exposes: a typed input struct, a typed output struct, a validation function, a calculation function, unit tests, and reference-value tests. Put this in src/methods/<category>/<method>.rs.
• Use f64, enums for finite option sets (e.g. Alternative), and Result<T, error::Error> for anything that can fail. Never return a raw string from core calculation logic — construct an error::Error variant.
• Validate every input before calculating: alpha in (0,1), power in (0,1), standard deviations/variances positive, allocation ratios positive, dropout rates in [0,1). Reuse validation:: helpers instead of re-deriving these checks per method.
• Sample sizes round up to the next integer. Never hide a rounding decision — expose the raw estimate (if useful), rounded group sizes, total N, and achieved power recalculated after rounding.
• Numerical/root-finding methods need a documented convergence tolerance and max-iteration bound, and must return Error::ConvergenceFailure instead of looping unboundedly or panicking.
• statrs has no noncentral t/F/chi-square support — exact power methods that need these must be implemented in distributions/, validated against R's pt/pf/pchisq with ncp set (see handbook/references/bibliography.md).
• Tauri commands (apps/desktop/src-tauri/src/lib.rs) are thin: deserialize, call into clinsize-core, map Error to a UI-friendly message. No calculation logic in a #[tauri::command] function, ever.
• Every fixed statistical or numerical defect gets a regression test that fails on the old behavior first (red-green), kept permanently.