STATUS: item 5 global slice (optimize v0.6.0), go-simd v0.7.0

STATUS.md

@@ -1,6 +1,6 @@
 # Status
 
-As of 2026-10-31. "Released" means a tagged version; "merged" means on main without a tag yet. Every child organisation has a landing, a documentation site, a brand and a profile; [go-scikit/docs](https://go-scikit.github.io/docs/latest/) links them all.
+As of 2026-11-02. "Released" means a tagged version; "merged" means on main without a tag yet. Every child organisation has a landing, a documentation site, a brand and a profile; [go-scikit/docs](https://go-scikit.github.io/docs/latest/) links them all.
 
 | # | item | home | state | evidence |
 |---|---|---|---|---|
@@ -11,6 +11,7 @@ As of 2026-10-31. "Released" means a tagged version; "merged" means on main with
 | 5 (SLSQP slice) | constrained NLP: SLSQP (`Minimize(…, SLSQP, …)` with `Constraint`, `NonlinearConstraint`, `LinearConstraint`) | [go-numerics/optimize](https://github.com/go-numerics/optimize) ([docs](https://go-numerics.github.io/docs/latest/optimize/)) | v0.3.0 released; Hock–Schittkowski cases 20–94× faster than SciPy, n = 50 with 100 bounds 1.4× ([docs/perf.md](https://github.com/go-numerics/optimize/blob/main/docs/perf.md)); v0.3.1: oracle and benchmark harnesses tested in CI | 100/100 statuses and 88/100 runs on SciPy 1.18.1's exact path (nit/nfev/njev, multipliers) over 36 Hock–Schittkowski problems and 10 stress cases; the other 12 are runs where SciPy's own arm64 and x86_64 wheels disagree; 100% statement coverage, fuzzed |
 | 5 (interior-point slice) | constrained NLP: trust-region interior point (`Minimize(…, TrustConstr, …)`, SciPy `trust-constr`, fmincon `'interior-point'`), same constraint types | [go-numerics/optimize](https://github.com/go-numerics/optimize) ([docs](https://go-numerics.github.io/docs/latest/optimize/)) | v0.4.0 released; Hock–Schittkowski cases 65–130× faster than SciPy, n = 50 with 100 bounds 2.7× ([docs/perf.md](https://github.com/go-numerics/optimize/blob/main/docs/perf.md)); global methods still planned | on the same 100 runs: success/failure as SciPy 1.18.1 on 100/100, same status 95/100, objective within 1e-5 of SciPy or 1e-4 of the optimum; 59–68 runs on SciPy's exact path where SciPy's own two wheels disagree on 35; 100% statement coverage, fuzzed |
 | 5 (QP slice) | convex quadratic programming: `Quadprog` (Goldfarb–Idnani dual active set and OSQP v1.0.0's ADMM with Ruiz scaling, adaptive ρ, polishing and certificates; `QPSolver` updates and warm starts), MATLAB `quadprog`, OSQP | [go-numerics/optimize](https://github.com/go-numerics/optimize) ([docs](https://go-numerics.github.io/docs/latest/optimize/)) | v0.5.0 released; ADMM runs OSQP v1.0.0's iterations on 25 of 25 generated instances, in 0.5–0.65× its time on dense n ≥ 500 and 0.8–0.9× on the larger portfolios, but 1.04–1.49× in the solve phase and 2.2–4.9× in the setup on small sparse problems (go-sparse's AMD, symbolic analysis and LDLᵀ solves: open defect there); the active set 0.68–0.76× the `quadprog` package from n = 500, 2.0× at n = 50 (open defect) ([docs/perf.md](https://github.com/go-numerics/optimize/blob/main/docs/perf.md)); `QPAuto` threshold n + m = 2500 measured | status, iteration and ρ-update counts equal OSQP v1.0.0's on 38 of 38 planted and 12 of 12 sparse cases; 33 Maros–Mészáros problems (tiny and small tiers, pinned and checksummed): Clarabel's objective on 27, OSQP's own status on the other 6, active set on all 13 positive definite; 100% statement coverage, fuzzed |
-| 5 (rest)–22 | see [PLAN.md](PLAN.md) | | planned | |
+| 5 (global slice) | global optimisation: `Direct` (DIRECT and DIRECT-L, a port of SciPy 1.18.1's C DIRECT, MIT notice kept), `DifferentialEvolution` (12 strategies, constraints, integrality, polishing), `Basinhopping`, on a shared `RandomSource` reproducing NumPy's PCG64 and MT19937 streams | [go-numerics/optimize](https://github.com/go-numerics/optimize) ([docs](https://go-numerics.github.io/docs/latest/optimize/)) | v0.6.0 released; no measured row slower than SciPy ([docs/perf.md](https://github.com/go-numerics/optimize/blob/main/docs/perf.md)); coverage of scipy.optimize and the MATLAB toolboxes in [docs/coverage.md](https://github.com/go-numerics/optimize/blob/main/docs/coverage.md) | DIRECT: 618 / 618 arithmetic-only runs bit-identical to SciPy (x, fun, nfev, nit, first 25 points), same status and counts on the 152 transcendental runs; DE: 296 / 296 unpolished runs bit-identical (final population, energies, nfev, nit) for PCG64 and MT19937, constraints and integrality included; basin hopping: SciPy's accept decisions on 224 / 224 runs; 100% statement coverage, fuzzed |
+| 5 (rest)–22 | see [PLAN.md](PLAN.md); item 6 design in [go-statistics/stats docs/design-models.md](https://github.com/go-statistics/stats/blob/main/docs/design-models.md), home [go-statistics/models](https://github.com/go-statistics/models) created | | planned | |
 | — | embeddable Julia: oracle harness (M0), parser (M1) and lowering (M2), see [bib/julia.md](bib/julia.md); runtime design (M3) in [bib/julia-runtime.md](bib/julia-runtime.md) | [go-embedded-julia/julia](https://github.com/go-embedded-julia/julia) | v0.1.0 (oracle + lexer), v0.2.0 (complete JuliaSyntax port), v0.3.0 (M2 lowering: the femtolisp front end of julia 1.12.6, `lower.NewState().Lower(ex)` = `Meta.lower`), v0.4.0 (package `core`: types, `apply_type`, subtyping and method specificity) and v0.5.0 (package `core/dispatch`: method tables, world ages, dispatch of concrete calls) released; corpus parsed in 0.70× julia's `Meta.parseall` time on one core; subtyping 1.49× **slower** than julia's `jl_subtype` on one core (3.63 ms vs 2.44 ms for the 20 000 pairs; open defect, allocation-bound, a pooling branch awaits a low-load A/B); M3 (Core runtime) in progress: types, subtyping, specificity and dispatch done; type intersection, builtins, intrinsics and the interpreter next | parser: tokens, green trees and `Expr` identical to julia 1.12.6 on 1191 / 1191 files of its `share/julia` tree and 976 / 976 snippets, error messages included; where JuliaSyntax itself throws and julia falls back to flisp (3 fuzz-found inputs) an error is returned. Lowering (v0.3.0), identical to `Meta.lower` modulo gensym numbering: 53 997 / 53 997 corpus forms (1191 / 1191 files), 124 / 124 snippets, 48 / 48 hygienic-scope forms, 477 / 477 built expressions, `deparse` 119 / 119 and front-end functions 151 / 151 against `julia --lisp`; 100% statement coverage; known gap: a `Float64` in a lowering error message returns `lower.ErrUnsupported`. `core` (v0.4.0), a port of jltypes.c and subtype.c of 1.12.6 without intersection: subtyping identical to julia on 20 000 random pairs and 4 902 cases from `test/subtype.jl`; specificity (`jl_type_morespecific`) identical to `Base.morespecific` on the same corpus. `core/dispatch` (v0.5.0): the selected method, ambiguity and no-method answers and the static parameters identical to julia's `which` and `jl_type_intersection_with_env` on 9 000 calls over 1 500 random method tables; 100% statement coverage, mutation-tested |
-| — | vector-update kernels `Axpy`, `Xpby`, `AxpyDot`, `Scale` (shared by the solvers) | [go-simd/floats](https://github.com/go-simd/floats) | v0.4.0 released: SIMD on amd64 and arm64, FMA loop elsewhere | — |
+| — | vector-update kernels `Axpy`, `Xpby`, `AxpyDot`, `Scale` (shared by the solvers) | [go-simd/floats](https://github.com/go-simd/floats) | v0.7.0 released: SIMD on amd64 and arm64; fused kernels (`Axpy`, `Axpby3`) for tolerance-validated solvers, and unfused ones (`AxpyNoFMA`, `XpbyNoFMA`, `Axpby3NoFMA`, `AxpyDiffNoFMA`) plus `SumPairwise` (NumPy's 8 accumulators in the lanes) for solvers validated by equality | unfused kernels equal to NumPy on 13 088 / 13 088 cases and `SumPairwise` to `np.sum` on 6 990 / 6 990, on arm64 and amd64 |