fix(recall): default embed model to Qwen3-Embedding-8B

deploy: multi-backend (minicpm/nemotron/finetuned) + 3-tier ZeroGPU->serverless->local-finetune fallback; finetune pipeline

feat: similar-task recall via Nebius embeddings (keyless; recall degrades off without NEBIUS_API_KEY)

deploy: offgrid on-device backend

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload src/unstuck/store.py with huggingface_hub

Upload src/unstuck/prompts.py with huggingface_hub

Upload src/unstuck/store.py with huggingface_hub

Upload src/unstuck/prompts.py with huggingface_hub

fix: thread-safe SQLite store

fix: unpack BatchEncoding into model.generate

fix(zerogpu): load model with .to(cuda) instead of device_map

Deploy Unstuck (Codex-built) — initial code