fix model filename mismatch, lighter neg prompt

add MCP server, REST API, docstring, gradio[mcp], ~30 min timing

preload models into page cache, offload-to-cpu, IQ4_XS text enc, conv-direct, mmap, vae f32, miku theme

use 4-step distill Q5_0 GGUF from WeReCooking repo

revert to 8-step Q5_0 GGUF (4-step OOMs during build conversion)

fix: use q5_0 quantization type (q5_k_m not supported by sd.cpp)

4-step distill: download BF16, convert to Q5_K_M GGUF at build time

Z-Anime 6B CPU: distill 8-step Q5_0, Qwen3-4B Q8_0, euler_a, beta schedule