sgBLAS v0.1.0

First evidence-gated sgBLAS release.

On one NVIDIA A100-SXM4-80GB, the tagged shape-dispatched SM80 portfolio reached a six-process median of 17.736 TFLOP/s at 4096 x 4096 x 4096. Across the four declared large NN shapes, the geometric mean of per-shape median throughput ratios was 94.08% versus cuBLAS configured with CUBLAS_COMPUTE_32F_PEDANTIC and TF32 disabled.

The release bundle is bound to commit b26974f. It contains all 45 checksummed evidence artifacts, deterministic source and evidence archives, the verifier report, an SPDX SBOM, a release manifest, and SHA256SUMS. All 24 checked correctness cases and all seven required dispatch paths passed; Compute Sanitizer memcheck, racecheck, initcheck, and synccheck were clean.

Scope: these measurements apply only to the declared GPU, corpus, strict-FP32 contract, and timing protocol. They are not a claim of universal cuBLAS parity.