feat(snapshot): ship the embedding matrix int8-quantized

Stage a deterministic model2vec.q8 sidecar (129 MB fp32 -> 32.6 MB int8;
fp32 weights barely compress under zstd so the archive saving is
near-full) and drop model.onnx from the staged copy. Only the active
model2vec spec converts — feature-extraction models ship as-is. Falls back
to shipping fp32 on any conversion error.