liminis-context-graph v0.16.4 — edges point at the right entity, and paging for big graphs

[verveguy]

v0.16.4 finishes the extraction-quality work from 0.16.3, and makes large graphs practical to read over MCP. There's no storage migration and no schema change.

Re-ingesting to clean up a graph? Do it on 0.16.4. It combines 0.16.3's fixes (the identifier check, consolidated summaries and optional LLM-verified dedup) with this release's fixes for wrongly re-pointed edges and self-loops.

Upgrading

• Re-ingest graphs built by extraction on 0.16.3 or earlier. Endpoints that were re-pointed onto the wrong entity, and stored self-loops, aren't repaired in place.
• knowledge_get_episodes with no group now returns every group, not just liminis. Pass group_ids: ["liminis"] to keep the old scope. group_ids is now honoured; before, it was silently ignored.

Edges point at the right entity (#645)

When an extracted edge named an entity that wasn't in the current chunk, lcg fuzzy-matched it against that chunk's entities and rewrote it, even when the name existed exactly in the graph already. In a document family, a reference to an earlier document (ACME-ADR-24004) could be redirected onto whichever similar document the chunk mentioned. On the reporter's corpus that affected 221 endpoints, and 1.6% of edges became self-loops.

• An exact match in the graph now wins. Fuzzy matching only runs for names that don't exist anywhere.
• Self-loops are dropped. An edge whose two ends resolve to the same entity is counted in edges_dropped_self_loop and not stored.
• Short codes stay distinct. ACDS and ACDM are no longer merged. Acronyms and their expansions still merge.

Reading big graphs over MCP (#648)

knowledge_get_episodes and knowledge_list_entities used to return everything at once, sometimes over 600K characters, which overflowed MCP clients. Both now take:

• cursor: send "", then follow next_cursor until it's null. Paging is stable through inserts and deletes.
• fields: only the keys you ask for.
• name_prefix: a prefix filter.

Without these parameters, responses are unchanged. knowledge_status also no longer shows a misleading exists: false WAL block when a WAL root is configured.

What was verified

• CI: the full Rust suite, all six real-corpus end-to-end jobs, Windows with real index creation and queries, and release builds for all four targets with pinned extensions. No check was cancelled.
• Acceptance tests: dedicated tests for every behaviour above, including the reporter's sibling-document scenario as an integration test with zero re-pointed siblings and zero self-loops, and the Liminis app's existing get_episodes call shape.
• macOS: the sidecar-gated replay battery ran with no skips, and its timings are level with 0.16.3 when run alternately.
• Published artifacts: all four archives pass their checksums, every bundled extension matches its pinned hash, and the published binary exposes group_ids, cursor, fields and name_prefix over MCP.

Full detail: docs/releases/0.16.4.md · CHANGELOG