Code improvements (#21–#25)

[m96-chan]

Summary

• Optimize ternary_gemm weight unpacking (prefill path) #21: Optimize ternary_gemm weight unpacking — exploit TILE_K=32 matching I2_S group size to batch-unpack 4 elements per u32 read (8→2 global reads per thread, 4x reduction)
• Extract duplicated createUniform helper into shared utility #22: Extract createUniformBuffer helper — deduplicate 5 identical private methods across nn/ classes into a shared free function in buffer-pool.ts
• Add CI check for shader sync between TypeScript and Rust #23: Add shader sync CI check — diff all .wgsl files between TS and Rust directories to catch desynchronization
• Cache tokenizer reverse byte map (bytesToString perf) #24: Cache tokenizer reverse byte map — add static unicodeToByte cache so bytesToString() stops rebuilding a 256-entry Map on every token decode
• BufferPool: add size-class bucketing and eviction #25: BufferPool size-class bucketing — bucket by (usage, nextPow2(alignedSize)) for O(1) acquire instead of O(n) scan; add stats() and trim() methods

Test plan

• npm run build passes
• npm run lint passes
• cargo check --workspace passes
• Shader sync diff passes locally
• Load BitNet 2B-4T in web-chat, verify coherent output (regression for Optimize ternary_gemm weight unpacking (prefill path) #21)
• Load Falcon-E 1B, verify coherent output

🤖 Generated with Claude Code