Roadmap and known gaps

The scipy surface scijit does not cover yet. Most gaps are future work, a few larger because scipy implements them in C or C++, and a few are deliberate choices.

Subpackages

This release ships interpolate, optimize and integrate. Planned for later releases: fft, linalg, stats, special, spatial, signal, ndimage, sparse (sparse.linalg, sparse.csgraph), cluster and constants.

Missing routines (future work)

Within the shipped subpackages, in roughly the order expected:

subpackage	routines
optimize	least_squares (TRF/dogbox), dual_annealing
integrate	solve_bvp, implicit Radau/BDF
interpolate	RBFInterpolator, griddata (nearest mode)

least_squares is the large one. dual_annealing is a loop over an existing minimizer. An implicit Radau/BDF would be the first prange-safe stiff solver in the package; the wrapped LSODA runs sequentially. Also future work: ODRPACK (scipy.odr), a Fortran wrap the build already handles, and KDTree/cKDTree, larger because a jitclass cannot reference its own type, so the tree has to be flattened into index arrays and every traversal rewritten.

Provided by numba and numpy

numba and numpy already run these inside @njit, so scijit does not duplicate them: np.linalg (solve, inv, det, eig, svd, qr, cholesky), np.interp, np.convolve, np.trapz, np.roots, np.gradient. numba’s native coverage grows release to release. As it does, more of this surface, including routines scijit ports today, becomes available natively and moves out of scope.

Backed by C or C++ in scipy

scipy implements these with C or C++ libraries. scijit wraps Fortran, so it does not wrap those libraries directly. The algorithms are published, so an @njit port or Fortran wrap is possible future work.

routine	scipy library
linprog, milp	HiGHS
direct	DIRECT
ConvexHull, Voronoi, Delaunay, griddata (linear/cubic)	Qhull
fmin_tnc, trust-krylov	TNC, trlib

griddata’s linear and cubic modes need Delaunay; only its nearest mode is reachable without Qhull.

Deliberate scope limits

Choices, not gaps.

• Randomized routines (basinhopping, differential_evolution) take an integer seed. They are reproducible per seed and independent across parallel calls, but do not return scipy’s values, because numba cannot call np.random.Generator. Deterministic routines return scipy’s values.

• Some wrapped routines are sequential-only, because the Fortran keeps its callback in a module variable. The thread-safety table in compatibility.md says which.