fix: PATH A embed uses sentence-transformers directly, not LM Studio API

fix: PATH A embed model switched from nomic to BGE-base

fix: rank-based graph edges (top 2 + 0.40 floor) instead of threshold

fix: backfill only updates related_ids (not wipe claims/recall), raise BGE threshold to 0.62, README hackathon tags

refactor: abstract VL chat template into _make_vl_inputs, model-agnostic

fix: Qwen2.5-VL chat template, skip embed on empty summary

fix: VL device detection, failed-capture error state, HF_HOME=/data/.cache

fix: lazy VL import, bump transformers>=4.49

fix: embed device=cpu, VL model class, file input reset

fix: expose @spaces.GPU functions in app.py for ZeroGPU detection, drop bitsandbytes

feat: sync all features from local-guide — BRIEF synthesis, REVIEW, weekly thread, unified lm.py, new frontend build

fix: lower similarity threshold to 0.50 for MiniLM, backfill on every restart

fix: use BitsAndBytesConfig for 4-bit quant, add torchvision for vision model

feat: add backend, frontend, seed data, and HF Spaces wiring