feat: ruvllm-wasm v2.0.0 — first functional WASM publish

[ruvnet]

Summary

Published @ruvector/ruvllm-wasm@2.0.0 to npm with compiled WASM binaries, replacing the deprecated v0.1.0 placeholder.

What's Included

• KV Cache — Two-tier (FP32 tail + u8 quantized store) token management
• Memory Pooling — Arena allocator + buffer pool for minimal allocation overhead
• Chat Templates — Llama3, Mistral, Qwen, ChatML, Phi, Gemma format support
• HNSW Semantic Router — Bidirectional graph with cosine similarity
• MicroLoRA — Per-request adaptation (rank 1-4)
• SONA Instant — EMA quality tracking + adaptive rank
• Web Workers — Parallel inference with SharedArrayBuffer detection
• TypeScript — Complete .d.ts type definitions

Build Notes

Rust 1.91 has a WASM codegen bug (type mismatch in release profile). Workaround:

CARGO_PROFILE_RELEASE_CODEGEN_UNITS=256 CARGO_PROFILE_RELEASE_LTO=off \
  wasm-pack build crates/ruvllm-wasm --target web --scope ruvector --release

Not Yet Wired

• IntelligentLLMWasm (combined router+lora+sona) — commented out, needs API alignment
• WebGPU attention shader (matmul works, attention falls back to CPU)
• GGUF model streaming loader
• Worker pool proper task completion (uses setTimeout polling)

Related

• PR fix: resolve 5 P0 critical issues + pre-existing compile errors #239 — Deprecated @ruvector/ruvllm-wasm@0.1.0 placeholder
• ADR-084 — ruvllm-wasm publish documentation