Same sj546 encoder checkpoint already on main, re-quantized from the same
fp32 model.onnx with --per_channel --reduce_range.
Per-channel alone produced full-scale +-127 weights. On x86-64 with AVX2 but
no VNNI, ORT's MLAS kernel accumulates adjacent u8*s8 products into int16 via
VPMADDUBSW and 2*255*127 = 64770 silently saturates at 32767, corrupting the
embedding. reduce_range caps weights at +-64 (2*255*64 = 32640, exact).
Before: 577394/7091712 weights over +-64 (max |w| = 127)
After: 0/7091712 weights over +-64 (max |w| = 64)
Graph unchanged: 142 initializers, no name or shape differences.
Mirrors rolf-mozilla/form-autofill-embed v0.3.1.
Toolchain: transformers.js 3.8.1 scripts.quantize, onnxruntime 1.20.1, onnx 1.17.0.
