clic v0.1.0

clic — an open, vendor-neutral compute language. CUDA's grid-of-threads model without the lock-in; Apple Metal backend today, FPGA/custom silicon next.

Library (verified on GPU): GEMM (naive + tiled ~864 GFLOP/s), dense layers, attention, softmax, layernorm, conv2d, INT8 quant, reductions, activations, SHA-256 (~889 MH/s vs Apple CryptoKit), a triangle rasterizer.

Tooling: a clic CLI, a Python host API, and an end-to-end MLP inference demo — kernels compose into real inference. CI compile-gate on every kernel.

Hardware (clic-A1 / OSHI-A1): verified, synthesizable RTL (systolic array + a full GEMM accelerator block) that maps to DSPs and fits the ULX3S 85F; plus the full silicon doc set and the OSHI-A1 spec (131 TOPS INT8 target).

Honest note: performance is bounded by the target device — not "faster than CUDA." All silicon figures are v0.1 targets pending fabrication. See CHANGELOG.md.