fix(gpu): warm LLM into VRAM at startup + health-check grace period — stops the cold-load restart loop

fix(gpu): ollama v0.32.1 .tar.zst (correct URL) + zstd; extract to /usr; GPU CUDA libs baked

feat(gpu): Ollama LLM server (qwen2.5:3b) for GPU inference + on-demand T4 with 15min auto-sleep

fix(docker): keep wheel filename so pip accepts it

perf(llm): install prebuilt AVX2 llama-cpp wheel (no HF-build compile) + Qwen2.5-3B for citations

perf(llm): AVX2/FMA source build of llama-cpp + Qwen2.5-3B (reliable [n] citations) — utilizes the paid CPU's vector units, 3B follows citation instructions

feat(llm): local Qwen2.5-1.5B GGUF via llama-cpp (prebuilt CPU wheel, model baked) — real generative answers on the paid cpu-upgrade tier

fix(space): HF_HUB_OFFLINE=1 — baked models must load from cache, never network (prevents relaunch health timeouts)

feat(voice): NotebookLM-style voice UX

feat(voice): ASR small + Kokoro TTS (spoken replies); prebake models

fix(voice): install faster-whisper (free, torch-free) + prebake Whisper base; enable ASR

Install .[ingest,llm] so the huggingface LLM provider (openai SDK) loads

Deploy Auralynq RAG (Llama-3.3-70B via HF Inference Providers)