v0.11.1 — Faster recall on low-resource devices, same results

Headline

Recall stays fast on densely connected memories and on low-resource devices, with exactly the same results. The graph step no longer lists every path through the graph, so it takes about a second instead of up to minutes and writes next to no temporary files. The agent's database tools no longer hold up the rest of the API while they run, and reads no longer re-index stored text.

Fixed

• Recall's graph step produced one row per path from every seed, three hops deep, each carrying the entity's full text. It now keeps the best result per entity and depth and reads each entity once. Same rows, depths and scores; where two seeds reach an entity equally, the same one now wins every time. On a copy of a real memory, 25 recalls returned identical results in 71 s instead of 19 minutes, writing 23 MB of temporary files instead of 175 GB.
• recall_memory, quick_search, get_entity, view_tree and view_entity_relations now run in a worker thread, so a slow query no longer stalls /health or other requests.
• Recall's access counting re-tokenized every returned entity's text. The full-text trigger now fires only when the title, content or keywords are written (migration 006). The stored index is unchanged.

Upgrading from v0.11.0

Rebuild and restart with docker compose up -d --build; migration 006 runs at start. No data changes and nothing to re-embed.