Roadmap

Current implementation

The library implements 97 semantic operations and 404 type signatures across f32, f64, six integer types, masks and conversions. Bounded fused pipelines, resident storage and Node/browser Worker execution are available. Private ABI v6 and registry fingerprints keep JS and native/Wasm assets synchronized.

Current behavior is defined by the contracts and operation signatures. The verification archive retains dated results and their scope.

Remaining validation

Item	Required evidence
Kernel organization comparison	Compare handwritten core::arch, a small internal trait/macro layer and pulp for add, signed-zero/NaN-sensitive min, select and a reduction; record correctness, target coverage, compile time, size and throughput, or document a decision to defer the comparison
Remaining native targets	Execute macOS x64, Linux glibc arm64/x64 and Windows x64 checks; retain matching source, binary and log hashes; actually run AVX2
Firefox	Complete static-package, Vite and Worker tests in a working browser environment; previous failures occurred before page assertions
Final candidate	Re-run acceptance from the final source checkpoint and the exact combined tarball; old dirty-workspace reports are historical only
Tool coverage	Record where Miri or applicable sanitizers can check pure Rust and where real SIMD/FFI testing is required
Performance tracking	Accumulate comparable fixed-machine baselines and noise bounds before setting regression thresholds

The retained candidate report is failed because Firefox could not start. Other native targets were pending. Local successes and CI configuration do not close those gaps. More SIMD coverage is valuable, but declared scalar fallback is supported; all 404 signatures need not be vectorized.

Release decision

Version naming, stable API/maintenance commitments and npm/tag publication are release decisions. Use the candidate workflow to collect matching platform and package evidence. Its fixed stableReady: false field does not encode an API stability or maintenance commitment.

Future candidates

Consider these only after a concrete workload justifies the additional semantics and maintenance cost:

• Additional measured SIMD kernels for f64, integers, masks and advanced math.
• Explicitly named relaxed/reordered numerical modes with independent budgets.
• More conversion pairs or wide integer accumulators with defined output types.
• Terminal reductions in fused plans and other bounded execution extensions.
• Additional platforms, AVX-512 or SVE.
• Multidimensional tensors, arbitrary strides, GPU, autodiff, BLAS, FFT, convolution, complex numbers or f16.

These are not implemented promises or criteria for finishing the current candidate.