[SamG-Coder]

Add a browser encoder at webgpu/encode.html that runs CUDA device kernels through CUDA WebShader and hands exported assets directly to the existing WebGPU viewer. Users can select one to six material maps, fit one-to-four-channel latent planes, compare reconstructions, and download the standard NNTC JSON plus DDS files as a ZIP.

The GPU path covers filtered observation generation, streamed decoder normal equations, four-color latent coordinate descent and reconstruction. Host code handles source mip/PCA initialization, small dense solves, quantization/BC packing and file I/O. The CUDA WebShader submodule pins 896432c, which adds the reusable fitting operation in SamG-Coder/cuda-webshader#2.

Scope: this is a format-compatible browser implementation, not native encoder parity. It uses f32 GPU accumulation and a different optimizer, and does not port all native material options, global sparse solves or decoder-objective BC refinement. The native implementation remains available. docs/BROWSER_ENCODER.md explains supported features and numerical differences.

Validation:

• Six Node tests pass, including all 32 channel/storage layouts through the existing viewer parser.
• Seven hardware WebGPU cases pass, covering six distinct output maps, odd sizes, split BC, mip chains, constants, repeated output, early/mid-run cancellation and GPU buffer cleanup. Independent exported-file reconstruction agrees within 2.71e-7.
• Real UI tests encode the bundled 512x512 single-map and four-map materials, download ZIPs, load the interactive viewer and check desktop/mobile layouts with no browser errors.
• The upstream Python DDS decoder reads both ZIP exports successfully; its single-map result is 35.16 dB.
• Local Edge/WebGPU default measurements: approximately 2.8 s / 35.2 dB for m1 and 3.5-3.6 s for m1-m4. Not a native CUDA benchmark.
• Optional SwiftShader could not obtain an adapter in installed Edge and is not counted as a pass.

Run: git submodule update --init --recursive; npm ci; npm start. Open http://127.0.0.1:5174.