revert to working P=128 while a16 int16-calib is tuned

a16 P=512 prefill, 8-sample calib (better int16 activation scales)

P=512 a16 (int16 acts, FLOAT I/O, int8 per-channel weights): fix -130 fp16-accum overflow, host-op binds f32

P=512 prefill W8A16 (int16 acts + int32 accum): fix multi-turn -130; user-diagnosed fp16-accumulator overflow

revert to working P=128: P=512 monolithic prefill drifts on v79 (both toolchains, empty output); pursuing runtime fallback

P=512 prefill CANDIDATE (user-authorized on-device validation; revert if fails)

revert prefill to P=128: P=512 build (mixed 2.48/2.45 toolchain) produced empty output on v79; rebuilding consistently

prefill P=128->512: fix multi-turn -130 (prompt>128 tokens); w8a16, export-exact 7/7 vs HF gold

Upload v79/llama_full_wqo_o3.bin with huggingface_hub