[krystophny]

A default-real Cardano exponent can move a bracketed spline root beyond the cell endpoint, so spline_roots drops it before Newton refinement. Preserve r8 precision in the exponent.

Fixes #316.

Adds a native regression with an independent Hermite-bisection oracle, root count/position/residual checks, and three amplitude scales. The reproducer fails on upstream master 2938fac9; the fix passes there and on develop 3e194677 through make -C install check-spline-root-endpoint. The physical spline is unchanged. A fresh native PENTRC replay reproduces the old failure and completes repaired FGAR/TGAR/PGAR with finite NetCDF output; registered energy controls pass unchanged. Torque convergence remains separate.

Chris&AI

[matt-pharr]

nice simple one-line change. In reality we should probably have every decimal number represented as _r8 across the code... I did a pass last year at getting most of the ones that were important but clearly there are some remaining.

PS I really wish Claude would stop naming things after people that don't need names. Just saying r4 vs r8 exponent would have told the average reader a lot more :^)