🔎 BGE-M3 In-Browser Retrieval

Scope: in-browser inference produces BGE-M3's dense embedding (CLS pooling, normalized). Its sparse/lexical and ColBERT multi-vector signals need the Python FlagEmbedding library and a compute backend, so they're outside this static demo. First run downloads the quantized ONNX weights (Xenova/bge-m3); the browser caches them afterwards.