`src/numeric.rs` — internal safe ratio math (pub(crate))

The crate denies arithmetic_side_effects and precision-loss casts under clippy pedantic; every “turn two counts into a float ratio” in the codebase goes through this module to stay lint-clean and panic-free. Not public API.

Key items

• trait RatioInt — the integer families used (u64 → f64, usize → f32), with lossless narrowing types (u32, u16) and the checked/saturating operations the lint set requires.
• unit_ratio(numerator, denominator) -> Float — the one function everything calls.

Behavior notes

• Zero denominator returns 0.0 (never NaN/inf) — “empty set” semantics at every call site.
• Results are not clamped to [0,1] — overflow/utilization reporting legitimately produces >1.0.
• Large operands are scaled down through a lossless intermediate so precision survives (a small numerator over a huge denominator stays strictly positive; boundary values don’t collapse to 1.0).
• Used by: convergence similarity, health scores, context utilization, memory relevance.

Deep dive: not user-facing — this page is the whole story.