OLLAMA_URL default was 192.168.0.32, a host that has been retired -- a fresh
install would have come up unable to reach any engine. The LLM now runs on the
Docker host, reachable from a container via the bridge gateway.
VALIDATOR_MODEL default was gemma4:31b (21.8 GB), which does not fit this
box's 16 GB card: it ran at 2.4 tok/s with 46% of the model on CPU and timed
out the validator stage at 600 s. qwen3-vl:8b fits entirely in VRAM at
44 tok/s.
Defaults only -- no fields added or removed, no required-flag changes, so no
form_fields schema migration for existing installs.
alexz
alexz