kg-corelisted

# Knowledge Graph Reference ## Session start The preload block is a compact core — the top-scored slice of both graphs, not the whole. The one full `kg_read(session_id)` renders everything it dropped without repeating what is already shown (preloaded gists collapse to id-only anchors). The session_id from the preload — or from the first kg_read — goes on every later kg_* call: it keeps one session, includes the project graph in searches, and avoids minting spurious sessions. No preload block (Desktop sessions, server still warming up): `kg_read(cwd="<project root>")` returns the full graph plus your session_id. Connection refused usually means the server is starting — retry after a few seconds; persistent trouble is a /kg-ops matter. Resuming an earlier conversation: `kg_sync(session_id)` picks up what other sessions wrote meanwhile. ## Recall After the full read, scan gists for anything touching the task and read those nodes in depth — several per call: `kg_read(session_id, ids=[...])`. Lean toward reading more: a wrong guess costs one call, missing context costs the task. Node reads return the node's own edges — each a crumb to the next hop — and reading an archived node promotes it and surfaces its orphaned neighbours. Three tiers as a graph grows: **active** (id + gist on the surface), **archived** (id + edges as crumb trails), **orphaned** (invisible — only search reaches them). ### Searching below the surface What kg_read renders is the surface of the graph, no