llama-cpp in systemPackages was dumping libggml-*.so into /run/current-system/sw/bin, which a GLib/GIO module scan tripped over on shell startup (spurious g_io_module_load warnings). Drop it. Tear down the rest of the unused local-AI stack: - ollama: ryu (was enable=false, only kept the ollama.darksailor.dev vhost), plus the already-disabled tako/shiro/kuro/darwin-module configs - open-webui (tako llms.nix) + chat.darksailor.dev + glance tile - ollama provider refs in opencode.nix and the aichat ryu client - aichat itself + the prepare-commit-msg hook that depended on it Removes ollama.darksailor.dev and chat.darksailor.dev subDomains (octodns will prune on next tako deploy).