KB embedding model cache (offline-first builds)

Xenova/bge-small-en-v1.5 (384-dim, quantized ONNX) — the pipeline's mandatory KB embedder, served from our own release so hosted builds never depend on huggingface.co availability or its rate limits (GH Actions runner IPs get HTTP 429 from HF — observed live 2026-07-17, run 29557533994).

Upstream: BAAI/bge-small-en-v1.5 (MIT license) · ONNX conversion by Xenova (huggingface.co/Xenova/bge-small-en-v1.5). Redistributed unmodified under MIT with attribution. Verified before upload: loads via @xenova/transformers 2.17.2 and produces real 384-dim embeddings (kb/warm-model-cache.mjs).