fortran

Imported from zekkr/fast_llqr (src/fortran/AGENTS.md). Install upstream with npx skills add zekkr/fast_llqr --skill fortran. Copyright stays with the author.

Fortran Kernel Instructions

These instructions apply to files under src/fortran/.

The Fortran kernels implement performance-critical versions of the screened warm-start ppro and stable seq algorithms. Treat them as numerical kernels, not as wrappers. Do not replace a ppro kernel with a call or behavior equivalent to the seq implementation.

Required Behavior

• No silent seq replacement. A ppro routine must implement the screened ppro algorithm. Any seq fallback must happen through an explicit wrapper or explicit fallback branch that reports status.
• No silent acceptance of nonconvergence. If the simplex loop reaches maxit, fails to find a valid pivot, or detects invalid state, return an explicit error/status flag.
• No stale aggregate rows. Rebuild aggregate rows whenever weights, evaluation point, threshold, sl, sh, S, bad signs, or reduced problem dimensions change.
• Recompute counts after changing sl, sh, or S. Do not reuse old reduced dimensions, offsets, or aggregate indices after threshold doubling or bad-sign insertion.
• Return explicit error/status flags. Prefer existing ierr-style outputs where present; do not hide kernel failures behind plausible numeric output.
• Preserve R/Fortran interface signatures unless a task explicitly asks for an interface change. If an interface change is required, update the R wrapper and document the compatibility impact.

Validation

Fortran changes must be validated against both:

• the R reference implementation for the same algorithm path, where available;
• the stable seq solver's H_seq, comparing each evaluation row as a set rather than an ordered vector.

Fallback results must be reported separately from ppro kernel successes.