[SamG-Coder]

NNTC's browser encoder needs to fit a small affine model from many GPU-generated observation rows without reading all rows back to JavaScript. Add a reusable NormalEquations accumulator: it streams row batches through deterministic GPU partial sums and performs the final small, pivoted solve on the host.

The API validates buffer ownership, row layout and bounds, supports multiple output channels and a relative ridge, and rejects empty or singular fits. GPU arithmetic is f32; this does not claim native f64/QR conditioning or accuracy.

Validation:

• npm test: 785 tests passed.
• node scripts/test-least-squares.mjs: 1,031 rows streamed in two batches on Edge/WebGPU; maximum coefficient difference from the CPU reference was 1.19e-7, with no WebGPU errors.
• NNTC integration exercised repeated fits and buffer cleanup during normal completion and cancellation.