fix: replace 130072-line vocab with the original 130560-line vocab the weights were quantized against

pondernet: add difficulty-conditioned prior (--difficulty-prior, PonderLoader) + make_gsm8k_zh_data.py (HF dataset -> difficulty-bucketed CoT jsonl) + README section 10

add pondernet/: PonderNet loop-thinking for MiniCPM5-2B (impl + train + eval + data + T4-verified head), separate subdir to avoid touching existing files

v2: profile-guided CPU inference optimization

v2: profile-guided CPU inference optimization

add vocab.tsv

add convert_minicpm_mixed.py

add server/minicpm_server.c

add README.md

add minicpm_mixed.bin

initial commit